main

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,13 @@
+# Jupyter Notebook
+__pycache__/
+*.pyc
+.ipynb_checkpoints/
+*.ipynb_checkpoints/*
+.ipynb_checkpoints/*
+src/samples
+# cache
+cache
+datasets
+test
+wandb
+nohup.out

TRAIN.md

@@ -0,0 +1,44 @@
+---
+license: apache-2.0
+---
+
+Краткая инструкция по установке
+Обновите систему и установите git-lfs:
+
+```
+apt update
+apt install git-lfs
+git config --global credential.helper store
+```
+Обновите pip и установите требуемые пакеты:
+
+```
+python -m pip install --upgrade pip
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124 -U
+pip install flash-attn --no-build-isolation # optional
+```
+Клонируйте репозиторий:
+
+```
+git clone https://huggingface.co/AiArtLab/sdxs
+cd sdxs/
+pip install -r requirements.txt
+```
+Подготовьте датасет:
+
+```
+mkdir datasets
+cd datasets
+huggingface-cli download AiArtLab/384 --local-dir 384 --repo-type dataset
+```
+Выполните вход в сервисы:
+
+```
+huggingface-cli login
+wandb login
+```
+Запустите обучение!
+
+```
+nohup accelerate launch train.py &
+```

butterfly.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b923bef9a5d1fe7103e960c943c110ec46155fc71d7f45e0070f3ef072bbdcb
+size 237918081

requirements.txt

@@ -0,0 +1,11 @@
+# torch>=2.6.0
+# torchvision>=0.21.0
+# torchaudio>=2.6.0
+diffusers>=0.32.2
+accelerate>=1.5.2
+datasets>=3.5.0
+matplotlib>=3.10.1
+wandb>=0.19.8
+huggingface_hub>=0.29.3
+bitsandbytes>=0.45.4
+transformers

src/dataset_from_folder.py

@@ -0,0 +1,423 @@
+# 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
+from torchvision.transforms import Resize, ToTensor, Normalize, Compose, InterpolationMode, Lambda
+from transformers import AutoModel, AutoImageProcessor, AutoTokenizer
+from typing import Dict, List, Tuple, Optional, Any
+from PIL import Image
+from tqdm import tqdm
+from datetime import timedelta
+
+# ---------------- 1️⃣ Настройки ----------------
+dtype = torch.float16
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+batch_size = 5
+min_size = 192 #256 #192
+max_size = 384 #256 #384
+step = 64
+img_share = 1.0
+empty_share = 0.05
+limit = 0
+textemb_full = False
+# Основная процедура обработки
+folder_path = "/workspace/butterfly" #alchemist"
+save_path = "/workspace/sdxs3d/datasets/butterfly" #"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 = AutoencoderKLWan.from_pretrained("AiArtLab/simplevae",subfolder="wan16x_vae_nightly",torch_dtype=dtype).to(device).eval()
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", subfolder=None,torch_dtype=dtype).to(device).eval()
+    
+    #vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae",subfolder="simple_vae_nightly",torch_dtype=dtype).to(device).eval()
+    #vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-schnell",subfolder="vae",torch_dtype=dtype).to(device).eval()
+    #vae = AutoencoderKL.from_pretrained("/home/recoilme/sdxs/vae", variant="fp16",torch_dtype=dtype).to(device).eval()
+    model = AutoModel.from_pretrained("visheratin/mexma-siglip2", dtype=dtype, trust_remote_code=True, optimized=True).to(device).eval()
+    processor = AutoImageProcessor.from_pretrained("visheratin/mexma-siglip2", use_fast=True)
+    tokenizer = AutoTokenizer.from_pretrained("visheratin/mexma-siglip2")
+    return vae, model, processor, tokenizer
+
+vae, model, processor, tokenizer = 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 encode_images_batch(images, processor, model, empty_share=0.0):
+    """
+    images: список PIL.Image
+    processor: трансформер для препроцессинга изображений
+    model: vision encoder (например, CLIP или подобный)
+    empty_share: доля эмбеддингов, которые нужно обнулить
+    """
+    # Преобразуем весь батч сразу (вместо обхода по каждому изображению)
+    processed = processor(images=images, return_tensors="pt")
+    pixel_values = processed["pixel_values"].to(device, dtype)
+
+    with torch.inference_mode():
+        outputs = model.vision_model(pixel_values)
+        #hidden_states = outputs.last_hidden_state  # [B, seq_len, dim]
+        pooled = outputs.pooler_output            # [B, dim]
+
+        # Добавляем pooled embedding в конец sequence
+        #context = torch.cat([hidden_states, pooled.unsqueeze(1)], dim=1)  # [B, seq_len+1, dim]
+        context = pooled.unsqueeze(1)
+
+        # Добавляем нулевые эмбеддинги с вероятностью empty_share
+        if empty_share > 0:
+            batch_size = context.shape[0]
+            num_empty = int(batch_size * empty_share)
+            if num_empty > 0:
+                zero_embeddings = torch.zeros_like(context[:num_empty])
+                context[:num_empty] = zero_embeddings
+
+    # Преобразуем bfloat16 в float32 если нужно
+    if context.dtype == torch.bfloat16:
+        context = context.to(torch.float32)
+
+    return context.cpu().numpy()  # [B, seq_len+1, dim]
+
+
+def encode_texts_batch(texts, tokenizer, model):
+    with torch.inference_mode():
+        text_tokenized = tokenizer(texts, return_tensors="pt", padding="max_length",
+            max_length=512,
+            truncation=True).to(device)
+        text_embeddings = model.encode_texts(text_tokenized.input_ids, text_tokenized.attention_mask)
+    return text_embeddings.unsqueeze(1).cpu().numpy()
+
+def encode_texts_batch_full(texts, tokenizer, model):
+    with torch.inference_mode():
+        text_tokenized = tokenizer(texts, return_tensors="pt", padding="max_length",max_length=512,truncation=True).to(device)
+        features = model.text_model(
+            input_ids=text_tokenized.input_ids, attention_mask=text_tokenized.attention_mask
+        ).last_hidden_state
+        features_proj = model.text_projector(features)
+    return features_proj.cpu().numpy()
+
+def clean_label(label):
+    label = label.replace("Image 1", "").replace("Image 2", "").replace("Image 3", "").replace("Image 4", "")
+    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
+
+            if latents_mean!=None and latents_std!=None:
+                latents = (latents - torch.tensor(latents_mean, device=device, dtype=dtype).view(1, -1, 1, 1, 1)) / torch.tensor(latents_std, device=device, dtype=dtype).view(1, -1, 1, 1, 1)
+            #print(latents.ndim, latents.shape)
+            if latents.ndim==5:
+                latents = latents[:, :, 0, :, :]  # Убираем временную ось [B, C, H, W]
+        
+        latents_np = latents.to(dtype).cpu().numpy()
+
+        # Обрабатываем тексты
+        text_labels = [clean_label(text) for text in texts]
+        if random.random() < img_share:
+            embeddings = encode_images_batch(pil_images, processor, model)
+            text_labels = [f"img: {label}" for label in text_labels]
+        else:
+            model_prompts, text_labels = process_labels_for_guidance(text_labels, empty_share)                
+            if textemb_full:
+                embeddings = encode_texts_batch_full(model_prompts, tokenizer, model)
+            else:
+                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=50000, 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=100000, batch_size=batch_size)
+
+# 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}")

src/model_create.ipynb

@@ -0,0 +1,514 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "5212f806-14b4-4b5f-bcb4-09e36df3b7d9",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "test unet\n",
+      "Количество параметров: 1616742724\n",
+      "Output shape: torch.Size([1, 4, 60, 48])\n",
+      "UNet2DConditionModel(\n",
+      "  (conv_in): Conv2d(4, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "  (time_proj): Timesteps()\n",
+      "  (time_embedding): TimestepEmbedding(\n",
+      "    (linear_1): Linear(in_features=288, out_features=1152, bias=True)\n",
+      "    (act): SiLU()\n",
+      "    (linear_2): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "  )\n",
+      "  (down_blocks): ModuleList(\n",
+      "    (0): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(288, 288, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 576, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=576, out_features=576, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=576, out_features=4608, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2304, out_features=576, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=576, out_features=576, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(288, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(288, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(576, 576, 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, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-7): 8 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(576, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, 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, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-7): 8 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 2304, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(2304, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(2304, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1728, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1728, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1728, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(1152, 1152, 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, 576, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=576, out_features=576, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=576, out_features=4608, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2304, out_features=576, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=576, out_features=576, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1728, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1728, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1728, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1152, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1152, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 864, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(864, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(864, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(576, 576, 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, 864, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(864, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(864, 288, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1-2): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(576, 288, 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, 1152, eps=1e-06, affine=True)\n",
+      "        (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        (transformer_blocks): ModuleList(\n",
+      "          (0-7): 8 x BasicTransformerBlock(\n",
+      "            (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn1): Attention(\n",
+      "              (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn2): Attention(\n",
+      "              (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (ff): FeedForward(\n",
+      "              (net): ModuleList(\n",
+      "                (0): GEGLU(\n",
+      "                  (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                )\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "                (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "      )\n",
+      "    )\n",
+      "    (resnets): ModuleList(\n",
+      "      (0-1): 2 x ResnetBlock2D(\n",
+      "        (norm1): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "        (conv1): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        (norm2): GroupNorm(32, 1152, eps=1e-05, affine=True)\n",
+      "        (dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (nonlinearity): SiLU()\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (conv_norm_out): GroupNorm(32, 288, eps=1e-05, affine=True)\n",
+      "  (conv_act): SiLU()\n",
+      "  (conv_out): Conv2d(288, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      ")\n"
+     ]
+    }
+   ],
+   "source": [
+    "config_sdxs = {\n",
+    "    # === Основные размеры и каналы ===\n",
+    "    \"in_channels\": 4,               # Количество входных каналов (совместимость с VAE)\n",
+    "    \"out_channels\": 4,              # Количество выходных каналов (симметрично in_channels)          \n",
+    "\n",
+    "    # === Cross-Attention ===\n",
+    "    \"cross_attention_dim\": 1152,      # Размерность текстовых эмбеддингов\n",
+    "    \"use_linear_projection\": True,\n",
+    "    \"norm_num_groups\": 32,\n",
+    "    \n",
+    "    # === Архитектура блоков ===\n",
+    "    \"down_block_types\": [ # энкодер\n",
+    "        \"DownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        #\"CrossAttnDownBlock2D\",\n",
+    "    ],\n",
+    "    \"up_block_types\": [   # декодер\n",
+    "        #\"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"UpBlock2D\",\n",
+    "    ],\n",
+    "\n",
+    "    # === Конфигурация каналов ===\n",
+    "    \"block_out_channels\": [288, 576, 1152],\n",
+    "\n",
+    "    \"transformer_layers_per_block\": [1, 1, 8],\n",
+    "    \"attention_head_dim\": [6, 9, 18],\n",
+    "}\n",
+    "\n",
+    "def check_initialization(model):\n",
+    "    for name, param in model.named_parameters():\n",
+    "        if param.requires_grad:\n",
+    "            print(f\"{name}: mean={param.data.mean():.3f}, std={param.data.std():.3f}\")\n",
+    "\n",
+    "\n",
+    "if 1:\n",
+    "    checkpoint_path = \"/workspace/sdxs3d/unet\"#\"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",
+    "    #new_unet = UNet2DConditionModel().to(\"cuda\", dtype=torch.float16)\n",
+    "\n",
+    "    # После инициализации\n",
+    "    #check_initialization(new_unet)\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(1,4, 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, 1152).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) "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1cb4ff0f-36cc-43cf-86a4-aaab9f106725",
+   "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": "3.11.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

src/model_create48.ipynb

@@ -0,0 +1,633 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "5212f806-14b4-4b5f-bcb4-09e36df3b7d9",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "test unet\n",
+      "Количество параметров: 1956883440\n",
+      "Output shape: torch.Size([1, 48, 60, 48])\n",
+      "UNet2DConditionModel(\n",
+      "  (conv_in): Conv2d(48, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "  (time_proj): Timesteps()\n",
+      "  (time_embedding): TimestepEmbedding(\n",
+      "    (linear_1): Linear(in_features=288, out_features=1152, bias=True)\n",
+      "    (act): SiLU()\n",
+      "    (linear_2): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "  )\n",
+      "  (down_blocks): ModuleList(\n",
+      "    (0): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(288, 288, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(48, 576, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=576, out_features=576, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=576, out_features=4608, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2304, out_features=576, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=576, out_features=576, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(288, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(288, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(576, 576, 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(48, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(576, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(1152, 1152, 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(48, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-7): 8 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, 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(48, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-7): 8 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-2): 3 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 2304, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(2304, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(2304, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(1152, 1152, 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(48, 1152, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 2304, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(2304, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(2304, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 1728, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1728, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "          (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1728, 1152, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (2): CrossAttnUpBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-2): 3 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(48, 576, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=576, out_features=576, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0): BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=576, out_features=576, bias=False)\n",
+      "                (to_k): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_v): Linear(in_features=1152, out_features=576, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=576, out_features=576, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((576,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=576, out_features=4608, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2304, out_features=576, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=576, out_features=576, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 1728, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1728, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1728, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1152, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1152, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 864, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(864, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=576, bias=True)\n",
+      "          (norm2): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(864, 576, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(576, 576, 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(48, 864, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(864, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(864, 288, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1-2): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(48, 576, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(576, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=1152, out_features=288, bias=True)\n",
+      "          (norm2): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(576, 288, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (mid_block): UNetMidBlock2DCrossAttn(\n",
+      "    (attentions): ModuleList(\n",
+      "      (0): Transformer2DModel(\n",
+      "        (norm): GroupNorm(48, 1152, eps=1e-06, affine=True)\n",
+      "        (proj_in): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        (transformer_blocks): ModuleList(\n",
+      "          (0-7): 8 x BasicTransformerBlock(\n",
+      "            (norm1): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn1): Attention(\n",
+      "              (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm2): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn2): Attention(\n",
+      "              (to_q): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_k): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_v): Linear(in_features=1152, out_features=1152, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm3): LayerNorm((1152,), eps=1e-05, elementwise_affine=True)\n",
+      "            (ff): FeedForward(\n",
+      "              (net): ModuleList(\n",
+      "                (0): GEGLU(\n",
+      "                  (proj): Linear(in_features=1152, out_features=9216, bias=True)\n",
+      "                )\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "                (2): Linear(in_features=4608, out_features=1152, bias=True)\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (proj_out): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "      )\n",
+      "    )\n",
+      "    (resnets): ModuleList(\n",
+      "      (0-1): 2 x ResnetBlock2D(\n",
+      "        (norm1): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "        (conv1): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (time_emb_proj): Linear(in_features=1152, out_features=1152, bias=True)\n",
+      "        (norm2): GroupNorm(48, 1152, eps=1e-05, affine=True)\n",
+      "        (dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (nonlinearity): SiLU()\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (conv_norm_out): GroupNorm(48, 288, eps=1e-05, affine=True)\n",
+      "  (conv_act): SiLU()\n",
+      "  (conv_out): Conv2d(288, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      ")\n"
+     ]
+    }
+   ],
+   "source": [
+    "config_sdxs = {\n",
+    "    # === Основные размеры и каналы ===\n",
+    "    \"in_channels\": 48,               # Количество входных каналов (совместимость с VAE)\n",
+    "    \"out_channels\": 48,              # Количество выходных каналов (симметрично in_channels)          \n",
+    "\n",
+    "    # === Cross-Attention ===\n",
+    "    \"cross_attention_dim\": 1152,      # Размерность текстовых эмбеддингов\n",
+    "    \"use_linear_projection\": True,\n",
+    "    \"norm_num_groups\": 48,\n",
+    "    \n",
+    "    # === Архитектура блоков ===\n",
+    "    \"down_block_types\": [ # энкодер\n",
+    "        \"DownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "    ],\n",
+    "    \"up_block_types\": [   # декодер\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"UpBlock2D\",\n",
+    "    ],\n",
+    "\n",
+    "    # === Конфигурация каналов ===\n",
+    "    \"block_out_channels\": [288, 576, 1152, 1152],\n",
+    "\n",
+    "    \"transformer_layers_per_block\": [1, 1, 1, 8],\n",
+    "    \"attention_head_dim\": [6, 12, 24, 24],\n",
+    "}\n",
+    "\n",
+    "def check_initialization(model):\n",
+    "    for name, param in model.named_parameters():\n",
+    "        if param.requires_grad:\n",
+    "            print(f\"{name}: mean={param.data.mean():.3f}, std={param.data.std():.3f}\")\n",
+    "\n",
+    "\n",
+    "if 1:\n",
+    "    checkpoint_path = \"/workspace/sdxs3d/unet\"#\"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",
+    "    #new_unet = UNet2DConditionModel().to(\"cuda\", dtype=torch.float16)\n",
+    "\n",
+    "    # После инициализации\n",
+    "    #check_initialization(new_unet)\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, 48, 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, 1152).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) "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1cb4ff0f-36cc-43cf-86a4-aaab9f106725",
+   "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": "3.11.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

train.py

@@ -0,0 +1,825 @@
+import os
+import math
+import torch
+import numpy as np
+import matplotlib.pyplot as plt
+from torch.utils.data import DataLoader, Sampler
+from torch.utils.data.distributed import DistributedSampler
+from torch.optim.lr_scheduler import LambdaLR
+from collections import defaultdict
+from torch.optim.lr_scheduler import LambdaLR
+from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL, DDPMScheduler
+from accelerate import Accelerator
+from datasets import load_from_disk
+from tqdm import tqdm
+from PIL import Image,ImageOps
+import wandb
+import random
+import gc
+from accelerate.state import DistributedType
+from torch.distributed import broadcast_object_list
+from torch.utils.checkpoint import checkpoint
+from diffusers.models.attention_processor import AttnProcessor2_0
+from datetime import datetime
+import bitsandbytes as bnb
+import torch.nn.functional as F
+from collections import deque
+
+# --------------------------- Параметры ---------------------------
+ds_path = "/workspace/sdxs3d/datasets/butterfly"
+project = "unet"
+batch_size = 16
+base_learning_rate = 9e-5
+min_learning_rate = 1e-5
+num_epochs = 30
+# samples/save per epoch
+sample_interval_share = 1
+use_wandb = False
+save_model = True
+use_decay = True
+fbp = False # fused backward pass
+optimizer_type = "adam8bit"
+torch_compile = False
+unet_gradient = True
+clip_sample = False #Scheduler
+fixed_seed = False
+shuffle = True
+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 = 1.03
+warmup_percent = 0.01
+dispersive_temperature=0.5
+dispersive_weight= 0.05
+percentile_clipping = 99 # 8bit optim
+betta2 = 0.995
+eps = 1e-8
+clip_grad_norm = 1.0
+steps_offset = 0 # Scheduler
+limit = 0
+checkpoints_folder = ""
+mixed_precision = "no" #"fp16"
+gradient_accumulation_steps = 1
+accelerator = Accelerator(
+    mixed_precision=mixed_precision,
+    gradient_accumulation_steps=gradient_accumulation_steps
+)
+device = accelerator.device
+
+# Параметры для диффузии
+n_diffusion_steps = 50
+samples_to_generate = 12
+guidance_scale = 5
+
+# Папки для сохранения результатов
+generated_folder = "samples"
+os.makedirs(generated_folder, exist_ok=True)
+
+# Настройка seed для воспроизводимости
+current_date = datetime.now()
+seed = int(current_date.strftime("%Y%m%d"))
+if fixed_seed:
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+# --- Пропорции лоссов и окно медианного нормирования (КОЭФ., не значения) ---
+# CHANGED: добавлен huber и dispersive в пропорции, суммы = 1.0
+loss_ratios = {
+    "mse":   0.60,
+    "mae":   0.35,
+    "huber": 0.0,
+    "dispersive": 0.05,
+}
+median_coeff_steps = 128  # за сколько шагов считать медианные коэффициенты
+
+# --------------------------- Параметры LoRA ---------------------------
+lora_name = "" 
+lora_rank = 32
+lora_alpha = 64
+
+print("init")
+
+# ---------------------------  вспомогательные функции ---------------------------
+def sample_timesteps_bias(
+    batch_size: int,
+    progress: float,           # [0..1]
+    num_train_timesteps: int,  # обычно 1000
+    steps_offset: int = 0,
+    device=None,
+    mode: str = "beta",        # "beta", "uniform"
+) -> torch.Tensor:
+    """
+    Возвращает timesteps с разным bias:
+    - beta   : как раньше (сдвиг в начало или конец в зависимости от progress)
+    - normal : около середины (гауссовое распределение)
+    - uniform: равномерно по всем timestep’ам
+    """
+
+    max_idx = num_train_timesteps - 1 - steps_offset
+
+    if mode == "beta":
+        alpha = 1.0 + .5 * (1.0 - progress)
+        beta  = 1.0 + .5 * progress
+        samples = torch.distributions.Beta(alpha, beta).sample((batch_size,))
+
+    elif mode == "uniform":
+        samples = torch.rand(batch_size)
+
+    else:
+        raise ValueError(f"Unknown mode: {mode}")
+
+    timesteps = steps_offset + (samples * max_idx).long().to(device)
+    return timesteps
+
+
+# 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
+
+    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:
+                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)
+
+class AccelerateDispersiveLoss:
+    def __init__(self, accelerator, temperature=0.5, weight=0.5):
+        self.accelerator = accelerator
+        self.temperature = temperature
+        self.weight = weight
+        self.activations = []
+        self.hooks = []
+    
+    def register_hooks(self, model, target_layer="down_blocks.0"):
+        unwrapped_model = self.accelerator.unwrap_model(model)
+        print("=== Поиск слоев в unwrapped модели ===")
+        for name, module in unwrapped_model.named_modules():
+            if target_layer in name:
+                hook = module.register_forward_hook(self.hook_fn)
+                self.hooks.append(hook)
+                print(f"✅ Хук зарегистрирован на: {name}")
+                break
+    
+    def hook_fn(self, module, input, output):
+        if isinstance(output, tuple):
+            activation = output[0]
+        else:
+            activation = output
+        if len(activation.shape) > 2:
+            activation = activation.view(activation.shape[0], -1)
+        self.activations.append(activation.detach().clone())
+        
+    def compute_dispersive_loss(self):
+        if not self.activations:
+            return torch.tensor(0.0, requires_grad=True, device=device)
+        local_activations = self.activations[-1].float()
+        batch_size = local_activations.shape[0]
+        if batch_size < 2:
+            return torch.tensor(0.0, requires_grad=True, device=device)
+        sf = local_activations / torch.norm(local_activations, dim=1, keepdim=True)
+        distance = torch.nn.functional.pdist(sf.float(), p=2) ** 2
+        exp_neg_dist = torch.exp(-distance / self.temperature) + 1e-5
+        dispersive_loss = torch.log(torch.mean(exp_neg_dist))
+        return dispersive_loss
+    
+    def clear_activations(self):
+        self.activations.clear()
+    
+    def remove_hooks(self):
+        for hook in self.hooks:
+            hook.remove()
+        self.hooks.clear()
+
+
+# --------------------------- Инициализация WandB ---------------------------
+if use_wandb and accelerator.is_main_process:
+    wandb.init(project=project+lora_name, config={
+        "batch_size": batch_size,
+        "base_learning_rate": base_learning_rate,
+        "num_epochs": num_epochs,
+        "fbp": fbp,
+        "optimizer_type": optimizer_type,
+    })
+
+# Включение Flash Attention 2/SDPA
+torch.backends.cuda.enable_flash_sdp(True)
+# --------------------------- Инициализация Accelerator --------------------
+gen = torch.Generator(device=device)
+gen.manual_seed(seed)
+
+# --------------------------- Загрузка моделей ---------------------------
+# VAE загружается на CPU для экономии GPU-памяти (как в твоём оригинальном коде)
+#vae = AutoencoderKLWan.from_pretrained("vae", variant="fp16").to(device="cpu", dtype=torch.float16).eval()
+#vae = AutoencoderKLWan.from_pretrained(
+#        "AiArtLab/simplevae", subfolder="wan16x_vae_nightly",
+#        torch_dtype=dtype
+#    ).to(device="cpu").eval()
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", subfolder=None,torch_dtype=dtype).to(device).eval()
+
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+latents_mean = getattr(vae.config, "latents_mean", None)
+latents_std = getattr(vae.config, "latents_std", None)
+
+# DDPMScheduler с V_Prediction и Zero-SNR
+scheduler = DDPMScheduler(
+    num_train_timesteps=1000,
+    prediction_type="v_prediction",
+    rescale_betas_zero_snr=True,
+    clip_sample = clip_sample,
+    steps_offset = steps_offset
+)
+
+
+class DistributedResolutionBatchSampler(Sampler):
+    def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
+        self.dataset = dataset
+        self.batch_size = max(1, batch_size // num_replicas)
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.shuffle = shuffle
+        self.drop_last = drop_last
+        self.epoch = 0
+        
+        try:
+            widths = np.array(dataset["width"])
+            heights = np.array(dataset["height"])
+        except KeyError:
+            widths = np.zeros(len(dataset))
+            heights = np.zeros(len(dataset))
+        
+        self.size_keys = np.unique(np.stack([widths, heights], axis=1), axis=0)
+        self.size_groups = {}
+        for w, h in self.size_keys:
+            mask = (widths == w) & (heights == h)
+            self.size_groups[(w, h)] = np.where(mask)[0]
+            
+        self.group_num_batches = {}
+        total_batches = 0
+        for size, indices in self.size_groups.items():
+            num_full_batches = len(indices) // (self.batch_size * self.num_replicas)
+            self.group_num_batches[size] = num_full_batches
+            total_batches += num_full_batches
+            
+        self.num_batches = (total_batches // self.num_replicas) * self.num_replicas
+        
+    def __iter__(self):
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        all_batches = []
+        rng = np.random.RandomState(self.epoch)
+        
+        for size, indices in self.size_groups.items():
+            indices = indices.copy()
+            if self.shuffle:
+                rng.shuffle(indices)
+            num_full_batches = self.group_num_batches[size]
+            if num_full_batches == 0:
+                continue
+            valid_indices = indices[:num_full_batches * self.batch_size * self.num_replicas]
+            batches = valid_indices.reshape(-1, self.batch_size * self.num_replicas)
+            start_idx = self.rank * self.batch_size
+            end_idx = start_idx + self.batch_size
+            gpu_batches = batches[:, start_idx:end_idx]
+            all_batches.extend(gpu_batches)
+        
+        if self.shuffle:
+            rng.shuffle(all_batches)
+        accelerator.wait_for_everyone()
+        return iter(all_batches)
+
+    def __len__(self):
+        return self.num_batches 
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+# Функция для выборки фиксированных семплов по размерам
+def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
+    size_groups = defaultdict(list)
+    try:
+        widths = dataset["width"]
+        heights = dataset["height"]
+    except KeyError:
+        widths = [0] * len(dataset)
+        heights = [0] * len(dataset)
+    for i, (w, h) in enumerate(zip(widths, heights)):
+        size = (w, h)
+        size_groups[size].append(i)
+    
+    fixed_samples = {}
+    for size, indices in size_groups.items():
+        n_samples = min(samples_per_group, len(indices))
+        if len(size_groups)==1:
+            n_samples = samples_to_generate
+        if n_samples == 0:
+            continue
+        sample_indices = random.sample(indices, n_samples)
+        samples_data = [dataset[idx] for idx in sample_indices]
+        latents = torch.tensor(np.array([item["vae"] for item in samples_data])).to(device=device,dtype=dtype)
+        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data])).to(device,dtype=dtype)
+        texts = [item["text"] for item in samples_data]
+        fixed_samples[size] = (latents, embeddings, texts)
+    
+    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям")
+    return fixed_samples
+
+if limit > 0:
+    dataset = load_from_disk(ds_path).select(range(limit))
+else:
+    dataset = load_from_disk(ds_path)
+
+def collate_fn_simple(batch):
+    latents = torch.tensor(np.array([item["vae"] for item in batch])).to(device,dtype=dtype)
+    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch])).to(device,dtype=dtype)
+    return latents, embeddings
+
+batch_sampler = DistributedResolutionBatchSampler(
+        dataset=dataset,
+        batch_size=batch_size,
+        num_replicas=accelerator.num_processes,
+        rank=accelerator.process_index,
+        shuffle=shuffle
+    )
+
+dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
+print("Total samples",len(dataloader))
+dataloader = accelerator.prepare(dataloader)
+
+start_epoch = 0
+global_step = 0
+total_training_steps = (len(dataloader) * num_epochs)
+world_size = accelerator.state.num_processes
+
+# Опция загрузки модели из последнего чекпоинта (если существует)
+latest_checkpoint = os.path.join(checkpoints_folder, project)
+if os.path.isdir(latest_checkpoint):
+    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device,dtype=dtype)
+    if torch_compile:
+        print("compiling")
+        torch.set_float32_matmul_precision('high')
+        unet = torch.compile(unet)
+        print("compiling - ok")
+    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)
+
+    # Создаём hook для dispersive только если нужно
+    if loss_ratios.get("dispersive", 0) > 0:
+        dispersive_hook = AccelerateDispersiveLoss(
+            accelerator=accelerator,
+            temperature=dispersive_temperature,
+            weight=dispersive_weight
+        )
+else:
+    # FIX: если чекпоинта нет — прекращаем с понятной ошибкой (лучше, чем неожиданные NameError дальше)
+    raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}. Положи UNet чекпоинт в {latest_checkpoint} или укажи другой путь.")
+
+if lora_name:
+    print(f"--- Настройка LoRA через PEFT (Rank={lora_rank}, Alpha={lora_alpha}) ---")
+    from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
+    from peft.tuners.lora import LoraModel
+    import os
+    unet.requires_grad_(False)
+    print("Параметры базового UNet заморожены.")
+
+    lora_config = LoraConfig(
+        r=lora_rank,
+        lora_alpha=lora_alpha,
+        target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+    )
+    unet.add_adapter(lora_config)
+
+    from peft import get_peft_model
+    peft_unet = get_peft_model(unet, lora_config)
+    params_to_optimize = list(p for p in peft_unet.parameters() if p.requires_grad)
+
+    if accelerator.is_main_process:
+        lora_params_count = sum(p.numel() for p in params_to_optimize)
+        total_params_count = sum(p.numel() for p in unet.parameters())
+        print(f"Количество обучаемых параметров (LoRA): {lora_params_count:,}")
+        print(f"Общее количество параметров UNet: {total_params_count:,}")
+
+    lora_save_path = os.path.join("lora", lora_name)
+    os.makedirs(lora_save_path, exist_ok=True)
+
+    def save_lora_checkpoint(model):
+        if accelerator.is_main_process:
+            print(f"Сохраняем LoRA адаптеры  в {lora_save_path}")
+            from peft.utils.save_and_load import get_peft_model_state_dict
+            lora_state_dict = get_peft_model_state_dict(model)
+            torch.save(lora_state_dict, os.path.join(lora_save_path, "adapter_model.bin"))
+            model.peft_config["default"].save_pretrained(lora_save_path)
+            from diffusers import StableDiffusionXLPipeline
+            StableDiffusionXLPipeline.save_lora_weights(lora_save_path, lora_state_dict)
+
+# --------------------------- Оптимизатор ---------------------------
+if lora_name:
+    trainable_params = [p for p in unet.parameters() if p.requires_grad]
+else:
+    if fbp:
+        trainable_params = list(unet.parameters())
+
+def create_optimizer(name, params):
+    if name == "adam8bit":
+        return bnb.optim.AdamW8bit(
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01,
+            percentile_clipping=percentile_clipping
+        )
+    elif name == "adam":
+        return torch.optim.AdamW(
+            params, lr=base_learning_rate, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01
+        )
+    elif name == "lion8bit":
+        return bnb.optim.Lion8bit(
+            params, lr=base_learning_rate, betas=(0.9, 0.97), weight_decay=0.01,
+            percentile_clipping=percentile_clipping
+        )
+    elif name == "adafactor":
+        from transformers import Adafactor
+        return Adafactor(
+            params, lr=base_learning_rate, scale_parameter=True, relative_step=False,
+            warmup_init=False, eps=(1e-30, 1e-3), clip_threshold=1.0,
+            beta1=0.9, weight_decay=0.01
+        )
+    else:
+        raise ValueError(f"Unknown optimizer: {name}")
+
+if fbp:
+    optimizer_dict = {p: create_optimizer(optimizer_type, [p]) for p in trainable_params}
+    def optimizer_hook(param):
+        optimizer_dict[param].step()
+        optimizer_dict[param].zero_grad(set_to_none=True)
+    for param in trainable_params:
+        param.register_post_accumulate_grad_hook(optimizer_hook)
+    unet, optimizer = accelerator.prepare(unet, optimizer_dict)
+else:
+    optimizer = create_optimizer(optimizer_type, unet.parameters())
+    def lr_schedule(step):
+        x = step / (total_training_steps * world_size)
+        warmup = warmup_percent
+        if not use_decay:
+            return base_learning_rate
+        if x < warmup:
+            return min_learning_rate + (base_learning_rate - min_learning_rate) * (x / warmup)
+        decay_ratio = (x - warmup) / (1 - warmup)
+        return min_learning_rate + 0.5 * (base_learning_rate - min_learning_rate) * \
+               (1 + math.cos(math.pi * decay_ratio))
+    lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
+
+    num_params = sum(p.numel() for p in unet.parameters())
+    print(f"[rank {accelerator.process_index}] total params: {num_params}")
+    for name, param in unet.named_parameters():
+        if torch.isnan(param).any() or torch.isinf(param).any():
+            print(f"[rank {accelerator.process_index}] NaN/Inf in {name}")
+    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
+    
+# Регистрация хуков ПОСЛЕ prepare
+if loss_ratios.get("dispersive", 0) > 0:
+    dispersive_hook.register_hooks(unet, "down_blocks.2")
+
+# --------------------------- Фиксированные семплы для генерации ---------------------------
+fixed_samples = get_fixed_samples_by_resolution(dataset)
+
+@torch.compiler.disable()
+@torch.no_grad()
+def generate_and_save_samples(fixed_samples_cpu, step):
+    original_model = None
+    try:
+        original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
+        vae.to(device=device).eval()  # временно подгружаем VAE на GPU для декодинга
+
+        scheduler.set_timesteps(n_diffusion_steps)
+        
+        all_generated_images = []
+        all_captions = [] 
+        
+        for size, (sample_latents, sample_text_embeddings, sample_text) in fixed_samples_cpu.items():
+            width, height = size
+            sample_latents = sample_latents.to(dtype=dtype, device=device)
+            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
+
+            noise = torch.randn(
+                sample_latents.shape,
+                generator=gen,
+                device=device,
+                dtype=sample_latents.dtype
+            )
+            current_latents = noise.clone()
+            
+            if guidance_scale > 0:
+                empty_embeddings = torch.zeros_like(sample_text_embeddings, dtype=sample_text_embeddings.dtype, device=device)
+                text_embeddings_batch = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
+            else:
+                text_embeddings_batch = sample_text_embeddings
+            
+            for t in scheduler.timesteps:
+                t_batch = t.repeat(current_latents.shape[0]).to(device)
+                if guidance_scale > 0:
+                    latent_model_input = torch.cat([current_latents] * 2)
+                else:
+                    latent_model_input = current_latents
+                
+                latent_model_input_scaled = scheduler.scale_model_input(latent_model_input, t_batch)
+                noise_pred = original_model(latent_model_input_scaled, t_batch, text_embeddings_batch).sample               
+                
+                if guidance_scale > 0:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                current_latents = scheduler.step(noise_pred, t, current_latents).prev_sample
+
+            #print(current_latents.ndim, current_latents.shape)
+            #if current_latents.ndim == 4:
+            #    current_latents = current_latents.unsqueeze(2)
+            # Латент  в форме [B, C, T, H, W]
+            #print(current_latents.ndim, current_latents.shape)
+
+            # Параметры нормализации
+            latent_for_vae = current_latents.detach() * scaling_factor + shift_factor
+
+            if latents_mean!=None and latents_std!=None:
+                latent_for_vae = latent_for_vae * torch.tensor(latents_std, device=device, dtype=dtype).view(1, -1, 1, 1, 1) + torch.tensor(latents_mean, device=device, dtype=dtype).view(1, -1, 1, 1, 1)
+            
+            decoded = vae.decode(latent_for_vae.to(torch.float32)).sample
+            #decoded = decoded[:, :, 0, :, :]  # [3, H, W]
+            #print(decoded.ndim, decoded.shape)
+            
+            decoded_fp32 = decoded.to(torch.float32)
+            for img_idx, img_tensor in enumerate(decoded_fp32):
+
+                # Форма: [3, H, W] -> преобразуем в [H, W, 3]
+                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
+                img = img.transpose(1, 2, 0)  # Из [3, H, W] в [H, W, 3]
+    
+                #img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
+                if np.isnan(img).any():
+                    print("NaNs found, saving stopped! Step:", step)
+                pil_img = Image.fromarray((img * 255).astype("uint8"))
+                
+                max_w_overall = max(s[0] for s in fixed_samples_cpu.keys())
+                max_h_overall = max(s[1] for s in fixed_samples_cpu.keys())
+                max_w_overall = max(255, max_w_overall)
+                max_h_overall = max(255, max_h_overall)
+            
+                padded_img = ImageOps.pad(pil_img, (max_w_overall, max_h_overall), color='white')
+                all_generated_images.append(padded_img)
+
+                caption_text = sample_text[img_idx][:200] if img_idx < len(sample_text) else ""
+                all_captions.append(caption_text)
+                
+                sample_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
+                pil_img.save(sample_path, "JPEG", quality=96)
+        
+        if use_wandb and accelerator.is_main_process:
+            wandb_images = [
+                wandb.Image(img, caption=f"{all_captions[i]}")
+                for i, img in enumerate(all_generated_images)
+            ]
+            wandb.log({"generated_images": wandb_images, "global_step": step})
+    finally:
+        # вернуть VAE на CPU (как было в твоём коде)
+        vae.to("cpu")
+        for var in list(locals().keys()):
+            if isinstance(locals()[var], torch.Tensor):
+                del locals()[var]
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# --------------------------- Генерация сэмплов перед обучением ---------------------------
+if accelerator.is_main_process:
+    if save_model:
+        print("Генерация сэмплов до старта обучения...")
+        generate_and_save_samples(fixed_samples,0)
+accelerator.wait_for_everyone()
+
+# Модифицируем функцию сохранения модели для поддержки LoRA
+def save_checkpoint(unet,variant=""):
+    if accelerator.is_main_process:
+        if lora_name:
+            save_lora_checkpoint(unet)
+        else:
+            if variant!="":
+                accelerator.unwrap_model(unet.to(dtype=torch.float16)).save_pretrained(os.path.join(checkpoints_folder, f"{project}"),variant=variant)
+            else:
+                accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
+            unet = unet.to(dtype=dtype)
+
+def batch_pred_original_from_step(model_outputs, timesteps_tensor, noisy_latents, scheduler):
+    device = noisy_latents.device
+    dtype = noisy_latents.dtype
+
+    available_ts = scheduler.timesteps
+    if not isinstance(available_ts, torch.Tensor):
+        available_ts = torch.tensor(available_ts, device="cpu")
+    else:
+        available_ts = available_ts.cpu()
+
+    B = model_outputs.shape[0]
+    preds = []
+    for i in range(B):
+        t_i = int(timesteps_tensor[i].item())
+        diffs = torch.abs(available_ts - t_i)
+        idx = int(torch.argmin(diffs).item())
+        t_for_step = int(available_ts[idx].item())
+        model_out_i = model_outputs[i:i+1]
+        noisy_latent_i = noisy_latents[i:i+1]
+        step_out = scheduler.step(model_out_i, t_for_step, noisy_latent_i)
+        preds.append(step_out.pred_original_sample)
+
+    return torch.cat(preds, dim=0).to(device=device, 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 = 1.
+
+for epoch in range(start_epoch, start_epoch + num_epochs):
+    batch_losses = []
+    batch_tlosses = []
+    batch_grads = []
+    batch_sampler.set_epoch(epoch)
+    accelerator.wait_for_everyone()
+    unet.train()
+    print("epoch:",epoch)
+    for step, (latents, embeddings) in enumerate(dataloader):
+        with accelerator.accumulate(unet):
+            if save_model == False and step == 5 :
+                used_gb = torch.cuda.max_memory_allocated() / 1024**3
+                print(f"Шаг {step}: {used_gb:.2f} GB")
+
+            noise = torch.randn_like(latents, dtype=latents.dtype)
+
+            progress = global_step / max(1, total_training_steps - 1)
+            timesteps = sample_timesteps_bias(
+                batch_size=latents.shape[0],
+                progress=progress,
+                num_train_timesteps=scheduler.config.num_train_timesteps,
+                steps_offset=steps_offset,
+                device=device
+            )
+
+            noisy_latents = scheduler.add_noise(latents, noise, timesteps)
+
+            if loss_ratios.get("dispersive", 0) > 0:
+                dispersive_hook.clear_activations()
+
+            #print(latents.shape,embeddings.shape)
+            model_pred = unet(noisy_latents, timesteps, embeddings).sample
+            target_pred = scheduler.get_velocity(latents, noise, timesteps)
+
+            # === Losses ===
+            losses_dict = {}
+
+            mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
+            losses_dict["mse"] = mse_loss
+            losses_dict["mae"] = F.l1_loss(model_pred.float(), target_pred.float())
+
+            # CHANGED: Huber (smooth_l1) loss added
+            losses_dict["huber"] = F.smooth_l1_loss(model_pred.float(), target_pred.float())
+
+            # === Dispersive loss ===
+            if loss_ratios.get("dispersive", 0) > 0:
+                disp_raw = dispersive_hook.compute_dispersive_loss().to(device)  # может быть отрицательным
+                losses_dict["dispersive"] = dispersive_hook.weight * disp_raw
+            else:
+                losses_dict["dispersive"] = torch.tensor(0.0, device=device)
+
+            # === 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)
+
+            # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
+            batch_losses.append(mse_loss.detach().item())
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+                
+            # Backward
+            accelerator.backward(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)
+                    grad = float(grad_val)
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+
+            global_step += 1
+            progress_bar.update(1)
+
+            # Логируем метрики
+            if accelerator.is_main_process:
+                if fbp:
+                    current_lr = base_learning_rate
+                else:
+                    current_lr = lr_scheduler.get_last_lr()[0]
+                batch_tlosses.append(total_loss.detach().item())
+                batch_grads.append(grad)
+
+                # Логируем только активные лоссы (ratio>0)
+                active_keys = [k for k, v in loss_ratios.items() if v > 0]
+                log_data = {}
+                for k in active_keys:
+                    v = losses_dict.get(k, None)
+                    if v is None:
+                        continue
+                    log_data[f"loss/{k}"] = (v.item() if isinstance(v, torch.Tensor) else float(v))
+
+                log_data["loss/total"] = float(total_loss.item())
+                log_data["loss/lr"] = current_lr
+                for k, c in coeffs.items():
+                    log_data[f"coeff/{k}"] = float(c)
+                if use_wandb and accelerator.sync_gradients:
+                    wandb.log(log_data, step=global_step)
+
+                # Генерируем сэмплы с заданным интервалом
+                if global_step % sample_interval == 0:
+                    generate_and_save_samples(fixed_samples,global_step)
+                    last_n = sample_interval
+                    avg_loss = float(np.mean(batch_losses[-last_n:])) if len(batch_losses) > 0 else 0.0
+                    avg_tloss = float(np.mean(batch_tlosses[-last_n:])) if len(batch_tlosses) > 0 else 0.0
+                    avg_grad = float(np.mean(batch_grads[-last_n:])) if len(batch_grads) > 0 else 0.0
+                    print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, grad: {avg_grad:.6f}")
+
+                    if save_model:
+                        print("saving:",avg_loss < min_loss*save_barrier)
+                        if avg_loss < min_loss*save_barrier:
+                            min_loss = avg_loss
+                            save_checkpoint(unet)
+                    if use_wandb:
+                        avg_data = {}
+                        avg_data["avg/loss"] = avg_loss
+                        avg_data["avg/tloss"] = avg_tloss
+                        avg_data["avg/grad"] = avg_grad
+                        wandb.log(avg_data, step=global_step)
+
+    if accelerator.is_main_process:
+        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses)>0 else 0.0
+        print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
+        if use_wandb:
+            wandb.log({"epoch_loss": avg_epoch_loss, "epoch": epoch+1})
+
+# Завершение обучения - сохраняем финальную модель
+if loss_ratios.get("dispersive", 0) > 0:
+    dispersive_hook.remove_hooks()
+if accelerator.is_main_process:
+    print("Обучение завершено! Сохраняем финальную моде��ь...")
+    if save_model:
+        save_checkpoint(unet,"fp16")
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+    
+print("Готово!")