grid

grid.py

@@ -0,0 +1,99 @@
+import torch
+from diffusers import DiffusionPipeline
+from pipeline_sdxs import SdxsPipeline
+
+device = "cuda"
+pipe_id = "/workspace/sdxs-1b"
+pipeline = SdxsPipeline.from_pretrained(
+    pipe_id,
+).to(device=device, dtype=torch.float16) 
+print(pipeline)
+prompts = [
+"A digital illustration of a young male character with white hair, blue eyes, and a dark complexion. He is wearing a white jacket, brown pants, and a brown belt. His hands are in his pockets, and he is wearing a wristwatch. He has a pouch on his belt and is wearing glasses. The background is a simple"
+,"A young woman with striking blue eyes and pointed ears, adorned with a floral kimono and a tattoo. Her hair is styled in a braid, and she wears a pair of ears"
+,"In the center of a futuristic space station, a highly advanced N7 soldier stands at attention. The soldier, clad in a suit of power armor, is equipped with a helmet that features a visor and a rifle slung over the shoulder. The armor is a striking combination of gray and red"
+,"A muscular, topless male with tiger-like ears and a tail stands in a forest, holding a sword and wearing a blue outfit, gazing directly at the viewer"
+,"A young woman with striking features, a mix of black and gold hair floating around her head, vibrant yellow eyes half-closed, and a blue sweater contrasting the black background, framed in a simple yet striking composition."
+,'A fluffy domestic cat with piercing green eyes sits attentively in a sunlit room filled with natural light, its soft fur reflecting warm hues of orange through golden windows.'
+,"A fierce woman in black-and-white, wearing a spiked iron mask with sharp metallic spikes, sporting a high ponytail, her skin marked with battle-dirt, and eyes that reflect determination and strength. The mask's menacing aura contrasts with her intense expression, capturing her mysterious and intimidating nature."
+,"A close-up of an astronaut's helmet with frosted, opaque visor, reflecting space's cold, frozen texture, resting on the visor a butterfly with vibrant, intricately patterned wings, and distant stars' faint glow."
+,"A striking character with red eyes and black uniform wields a sword in a defensive stance, poised for battle amidst a stark white background with vibrant red accents."
+,"A bald woman with a tattooed upper body stands against a vibrant red background. She is adorned with a pair of dangling earrings and a dress with a dragon design"
+,"white cyborg knight riding cyber horse with wings, long white gown, holding scythe, skeleton horse, zombies cyberpunk armor, feathers"
+,"A radiant voluptuous woman in Arizona's Grand Canyon at twilight, her fiery ginger hair cascading over a halter top with intricate lace, plunging neckline, black thongs, silver rings, and stacked bangles, surrounded by deep purples and burnt oranges in the sky, her light-blue eyes glowing with wonder as she gazes at the luminous beauty of the night."
+,"A man with hair, a white suit and black scarf, anime-style elements, glowing pink horns, and cyberpunk manga art style, standing in a shadowy smoke-filled background, with glowing pink eyes and a side view portrait of his face, white skin, and ear piercings."
+,"There is a young male character standing against a vibrant, colorful graffiti wall. he is wearing a straw hat, a black jacket adorned with gold accents, and black shorts."
+,"A black BMW M3 sports car with black and yellow rims"
+,"A young girl in a flowing, vibrant dress, her glowing eyes capturing the warmth of the day, sits on a grassy field, surrounded by anime-style elements."
+,"A young woman with a messy bun and vibrant red hoodie stands ready to strike with a large axe, her intense gaze and dynamic pose evoking energy and anticipation."
+,"A small bluebird with a white breast and black stripe perches on a branch, facing right with a blurred background of trees and bushes."
+,"A stylized, anime-style illustration of a young male character with white hair, wearing a white suit and holding two guns. He is standing with his arms outstretched and a confident smile on his face."
+,"In the center of a white background, a fantasy creature with blue eyes and a white body stands on its hind legs. Its front legs are adorned with armor, and its tail is adorned with scales. The creature has a large, spiky crest on its head and sharp claws"
+,"In a serene garden, two young girls stand side by side, their youthful energy palpable. The girl on the left, adorned with a blue dress and a matching blue flower in her hair, gazes directly at the viewer, her eyes sparkling with curiosity."
+]
+prompts2 = ["Космонавт на мотоцикле"]
+
+negative_prompt="low quality, bad quality, photo, low details"
+negative_prompt2 = "worst quality, low quality, photo, low details, blurry, jpeg artifacts, unfinished, sketch, sepia, missing limb, text, bad anatomy, bad proportions, bad hands, missing fingers"
+negative_prompt = "bad quality grainy image with low details, incomplete text, despite numerous technical flaws and distorted figures"
+
+all_images = []
+# Обработка батчей с прогресс-баром
+for i, prompt in enumerate(prompts):
+    image = pipeline(
+        prompt = prompt,
+        negative_prompt = negative_prompt,
+        guidance_scale = 5,
+        width = 1024,
+        height = 1408,
+        seed = 0,
+        batch_size = 1,
+    )[0]
+    all_images.extend(image)
+import matplotlib.pyplot as plt
+import math
+import numpy as np
+import textwrap
+
+def display_image_grid(images, prompts, cols=3, save_path=None):
+    n = len(images)
+    if n == 0:
+        return
+        
+    rows = math.ceil(n / cols)
+
+    # constrained_layout вместо tight_layout лучше держит одинаковую высоту ячеек
+    fig, axes = plt.subplots(rows, cols, figsize=(cols * 4, rows * 4.5), constrained_layout=True)
+
+    if isinstance(axes, np.ndarray):
+        axes = axes.flatten()
+    elif isinstance(axes, plt.Axes):
+        axes = [axes]
+        
+    axes = list(axes)[:n]
+
+    for i, (img, prompt) in enumerate(zip(images, prompts)):
+        ax = axes[i]
+        ax.imshow(img)
+        ax.axis("off")
+        ax.set_aspect("equal")
+
+        text = (prompt[:200] + "…") if len(prompt) > 200 else prompt
+        
+        # Разбиваем на строки
+        lines = textwrap.wrap(text, width=35)
+        
+        while len(lines) < 4:
+            lines.append("")
+            
+        ax.set_title("\n".join(lines), fontsize=9, pad=8)
+        # -----------------------------------
+
+    if save_path:
+        # Для сохранения constrained_layout работает автоматически
+        plt.savefig(save_path, bbox_inches="tight", dpi=600, format="jpeg")
+
+    #plt.show()
+
+
+display_image_grid(all_images, prompts, save_path="media/result_grid.jpg")

test.py

@@ -0,0 +1,34 @@
+import torch
+from diffusers import DiffusionPipeline
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+pipe_id = "/workspace/sdxs-1b"
+pipe = DiffusionPipeline.from_pretrained(
+    pipe_id,
+    torch_dtype=dtype,
+    trust_remote_code=True
+).to(device)
+
+prompt="There is a young male character standing against a vibrant, colorful graffiti wall. he is wearing a  hat, a  jacket adorned with gold accents, and black shorts."
+#prompt = "A stylized digital detailed illustration of a melancholy female character with white hair and distinct pointed ears, seated cross-legged, wearing a  sleeveless top featuring a bold red and black  design on her thighs, paired with black thigh-high boots."
+#prompt = "A young woman with striking blue eyes and pointed ears, with a dragon at background, adorned with a kimono and a tattoo. Her hair is styled in a braid, and she wears a pair of ears"
+refined = pipe.refine_prompts(prompt)[0]
+negative_prompt = "worst quality, low quality, photo, low details, blurry, jpeg artifacts, unfinished, sketch, sepia, missing limb, text, bad anatomy, bad proportions, bad hands, missing fingers"
+
+print(refined)
+output = pipe(
+    prompt=refined,
+    negative_prompt=negative_prompt,
+    #seed=43,
+)
+
+# Берем нулевую картинку и сам промпт (он будет строкой, так как мы подавали строку)
+image = output.images[0]
+refined_prompt_text = output.prompt 
+
+image.show()
+image.save("media/girl1.jpg")
+
+print(f"✨ refined_prompt_text:\n{refined_prompt_text}")

train4080.py

@@ -0,0 +1,837 @@
+import os
+import math
+import torch
+import numpy as np
+import matplotlib.pyplot as plt
+import wandb,comet_ml
+import random,time
+import gc
+import bitsandbytes as bnb
+import torch.nn.functional as F
+import argparse
+
+from diffusers import UNet2DConditionModel, AsymmetricAutoencoderKL, FlowMatchEulerDiscreteScheduler
+from transformers import Qwen3_5Tokenizer, Qwen3_5ForConditionalGeneration
+from torch.utils.data import DataLoader, Sampler
+from torch.optim.lr_scheduler import LambdaLR
+from collections import defaultdict
+from accelerate import Accelerator
+from datasets import load_from_disk
+from tqdm import tqdm
+from PIL import Image, ImageOps
+from torch.utils.checkpoint import checkpoint
+from diffusers.models.attention_processor import AttnProcessor2_0
+from datetime import datetime
+
+# Muon not tested! pip install git+https://github.com/recoilme/muon_adamw8bit.git
+from muon_adamw8bit import MuonAdamW8bit
+from transformers.optimization import Adafactor
+
+os.environ["NCCL_P2P_DISABLE"] = "1"
+os.environ["NCCL_IB_DISABLE"] = "1" # comment this on H100!
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+
+# --------------------------- Параметры ---------------------------
+ds_path = "datasets/alchemist_704_vae8x16x_imgpool"
+project = "unet"
+
+gpu_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
+local_bs = max(1, int((gpu_mem_gb / 32) * 7))
+num_gpus = torch.cuda.device_count()
+batch_size = local_bs * num_gpus
+
+base_learning_rate = 4e-5
+min_learning_rate = 4e-6
+
+# 0.5 - pretrain (base forms)
+# 1 - base train (composition)
+# 3 - finetuning (anatomy)
+# 5 - small details (faces)
+learning_rate_scale = 3
+base_learning_rate = base_learning_rate / learning_rate_scale
+min_learning_rate = min_learning_rate / learning_rate_scale
+print(f"Calculated params max-lr:{base_learning_rate} min-lr:{min_learning_rate} GPUs: {num_gpus}, Global BS: {batch_size}")
+
+num_epochs = num_gpus
+sink_interval_share = 20
+cfg_dropout = 0.10
+max_length = 248
+use_precomputed_embeddings = True
+use_wandb = False
+use_comet_ml = True
+save_model = True
+use_decay = True
+fbp = True
+torch_compile = False
+unet_gradient = True
+loss_normalize = False
+fixed_seed = False
+shuffle = True
+optimizer_type = "adam8bit"
+if optimizer_type == "muon_adam8bit":
+    batch_size = num_gpus * max(1, int((gpu_mem_gb / 32) * 3))
+    muon_lr_scale = 500
+comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r" # hardcoded for blind run, i don't care about key
+comet_ml_workspace = "recoilme" 
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+# MAX_JOBS=4 pip install flash-attn --no-build-isolation
+torch.backends.cuda.enable_flash_sdp(True)
+torch.backends.cuda.enable_mem_efficient_sdp(True)
+torch.backends.cuda.enable_math_sdp(False) # Отключаем медленный вариант
+save_barrier = 1.25
+warmup_percent = 0.0025
+betta2 = 0.997
+eps = 1e-7
+clip_grad_norm = 1.0
+limit = 0
+checkpoints_folder = ""
+gradient_accumulation_steps = 1
+dtype = torch.float32
+mixed_precision = "bf16"
+
+# Параметры для диффузии
+n_diffusion_steps = 40
+samples_to_generate = 12
+guidance_scale = 1
+
+# Папки для сохранения результатов
+generated_folder = "samples"
+os.makedirs(generated_folder, exist_ok=True)
+
+# Настройка seed
+current_date = datetime.now()
+seed = int(current_date.strftime("%Y%m%d")) + 42
+if fixed_seed:
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+accelerator = Accelerator(
+    mixed_precision=mixed_precision,
+    gradient_accumulation_steps=gradient_accumulation_steps
+)
+device = accelerator.device
+
+print("init")
+# Создаём объект ArgumentParser с рассчитанными значениями по умолчанию
+parser = argparse.ArgumentParser(description='Train a model on a dataset.')
+parser.add_argument('--ds-path', type=str, default=ds_path, help='Path to the dataset')
+parser.add_argument('--ep', type=int, default=num_epochs, help='Number of epochs to train the model')
+parser.add_argument('--batch', type=int, default=batch_size, help='Total batch size')
+parser.add_argument('--min-lr', type=float, default=min_learning_rate, help='Minimum learning rate')
+parser.add_argument('--max-lr', type=float, default=base_learning_rate, help='Maximum learning rate')
+parser.add_argument('--dry-run', action='store_true',default=False, help='Dry run train without saving/sampling')
+parser.add_argument('--lvl', type=float, default=0.0, help='Train level, from 0.5 to 5')
+
+# Парсим аргументы командной строки
+args = parser.parse_args()
+
+# Используем значения из аргументов
+batch_size = args.batch
+ds_path = args.ds_path
+base_learning_rate = args.max_lr
+min_learning_rate = args.min_lr
+num_epochs = args.ep
+lvl = args.lvl
+if args.dry_run:
+    save_model = False
+if lvl >= 0.1:
+    base_learning_rate = base_learning_rate / lvl
+    min_learning_rate = min_learning_rate / lvl
+    print(f"max-lr:{base_learning_rate} min-lr:{min_learning_rate}")
+    
+
+# --------------------------- Инициализация WandB ---------------------------
+if accelerator.is_main_process:
+    if use_wandb:
+        wandb.init(project=project, config={
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "num_epochs": num_epochs,
+            "optimizer_type": optimizer_type,
+        })
+    if use_comet_ml:
+        from comet_ml import Experiment
+        comet_experiment = Experiment(
+            api_key=comet_ml_api_key,
+            project_name=project,
+            workspace=comet_ml_workspace
+        )
+        hyper_params = {
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "num_epochs": num_epochs,
+        }
+        comet_experiment.log_parameters(hyper_params)
+
+# --------------------------- Загрузка моделей ---------------------------
+vae = AsymmetricAutoencoderKL.from_pretrained("vae", torch_dtype=dtype).to("cpu").eval()
+scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
+tokenizer = None
+text_encoder = None
+
+def load_text_encoder():
+    global tokenizer, text_encoder
+    if tokenizer is None:
+        tokenizer = Qwen3_5Tokenizer.from_pretrained("tokenizer")
+    if text_encoder is None:
+        text_encoder = Qwen3_5ForConditionalGeneration.from_pretrained(
+            "text_encoder",
+            torch_dtype=torch.float16
+        ).to(device).eval()
+        
+load_text_encoder()
+
+@torch.no_grad()
+def encode_texts(text, max_length=max_length):
+    if text is None: text = ""
+    if isinstance(text, str): text = [text]
+
+    formatted_prompts = []
+    for t in text:
+        messages = [{"role": "user", "content": [{"type": "text", "text": t}]}]
+        formatted_prompts.append(tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False))
+
+    toks = tokenizer(formatted_prompts, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt").to(device)
+    outputs = text_encoder(input_ids=toks.input_ids, attention_mask=toks.attention_mask, output_hidden_states=True)
+    
+    last_hidden = outputs.hidden_states[-2]
+
+    return last_hidden.to(dtype=torch.float16), toks.attention_mask.to(dtype=torch.bool)
+        
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+mean = getattr(vae.config, "latents_mean", None)
+std = getattr(vae.config, "latents_std", None)
+if mean is not None and std is not None:
+    latents_std = torch.tensor(std, device=device, dtype=dtype).view(1, len(std), 1, 1)
+    latents_mean = torch.tensor(mean, device=device, dtype=dtype).view(1, len(mean), 1, 1)
+
+import numpy as np
+from torch.utils.data import Sampler
+
+
+class DistributedResolutionBatchSampler(Sampler):
+    def __init__(self, dataset, batch_size, num_replicas, rank, drop_last=True, shuffle=True):
+        self.dataset = dataset
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.shuffle = shuffle
+        self.drop_last = drop_last
+        self.epoch = 0
+
+        # batch на одну GPU
+        self.batch_size = max(1, batch_size // num_replicas)
+        self.global_batch = self.batch_size * num_replicas
+
+        try:
+            widths = np.asarray(dataset["width"])
+            heights = np.asarray(dataset["height"])
+        except KeyError:
+            widths = np.zeros(len(dataset))
+            heights = np.zeros(len(dataset))
+
+        # --- группировка индексов ---
+        groups = {}
+        for i, (w, h) in enumerate(zip(widths, heights)):
+            groups.setdefault((w, h), []).append(i)
+
+        # --- создаём список всех глобальных батчей ---
+        all_batches = []
+
+        for indices in groups.values():
+
+            idx = np.asarray(indices, dtype=np.int64)
+
+            num_batches = len(idx) // self.global_batch
+            if num_batches == 0:
+                continue
+
+            idx = idx[: num_batches * self.global_batch]
+
+            batches = idx.reshape(num_batches, self.global_batch)
+
+            all_batches.append(batches)
+
+        if len(all_batches) > 0:
+            self.global_batches = np.concatenate(all_batches, axis=0)
+        else:
+            self.global_batches = np.empty((0, self.global_batch), dtype=np.int64)
+
+        self.num_batches = len(self.global_batches)
+
+    def __iter__(self):
+
+        rng = np.random.RandomState(self.epoch)
+
+        order = np.arange(self.num_batches)
+
+        if self.shuffle:
+            rng.shuffle(order)
+
+        start = self.rank * self.batch_size
+        end = start + self.batch_size
+
+        for i in order:
+            yield self.global_batches[i][start:end]
+
+    def __len__(self):
+        return self.num_batches
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+        
+
+
+# --- [UPDATED] Функция для фиксированных семплов (ХРАНИМ НА CPU) ---
+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]
+        
+        # Латенты лежат строго на CPU (без .to(device))
+        latents = torch.tensor(np.array([item["vae"] for item in samples_data]), dtype=dtype)
+        texts = [item["text"] for item in samples_data]
+        
+        if use_precomputed_embeddings:
+            # Эмбеддинги лежат строго на CPU
+            embeddings = torch.tensor(
+                np.array([item["embeddings"] for item in samples_data]),
+                dtype=torch.float16
+            )
+            masks = torch.tensor(
+                np.array([item["attention_mask"] for item in samples_data]),
+                dtype=torch.bool
+            )
+        else:
+            # encode_texts возвращает GPU тензоры, поэтому принудительно возвращаем их на CPU для хранения
+            embeddings, masks = encode_texts(texts)
+            embeddings = embeddings.cpu()
+            masks = masks.cpu()
+        
+        fixed_samples[size] = (latents, embeddings, masks, texts)
+    
+    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям (хранятся в RAM на CPU)")
+    return fixed_samples
+
+if limit > 0:
+    dataset = load_from_disk(ds_path).select(range(limit))
+else:
+    dataset = load_from_disk(ds_path)
+
+
+print(f"images: {len(dataset)}")
+
+def collate_fn_simple(batch):
+
+    latents = torch.from_numpy(
+        np.array([item["vae"] for item in batch], dtype=np.float16)
+    ).to(device, dtype=torch.float16)
+
+    if use_precomputed_embeddings:
+        embeddings = torch.from_numpy(
+            np.array([item["embeddings"] for item in batch], dtype=np.float16)
+        ).to(device, dtype=torch.float16)
+
+        attention_mask = torch.from_numpy(
+            np.array([item["attention_mask"] for item in batch], dtype=np.int64)
+        ).to(device)
+
+        return latents, embeddings, attention_mask
+
+    raw_texts = [item["text"] for item in batch]
+
+    texts = [
+        "" if t.lower().startswith("zero") 
+        else "" if random.random() < cfg_dropout
+        else t[1:].lstrip() if t.startswith(".")
+        else t.replace("The image shows ", "").replace("The image is ", "").replace("This image captures ","").strip()
+        for t in raw_texts
+    ]
+
+    embeddings, attention_mask = encode_texts(texts)
+    attention_mask = attention_mask.to(dtype=torch.bool)
+
+    return latents, embeddings, attention_mask
+
+batch_sampler = DistributedResolutionBatchSampler(
+        dataset=dataset,
+        batch_size=batch_size,
+        num_replicas=accelerator.num_processes,
+        rank=accelerator.process_index,
+        shuffle = shuffle
+    )
+
+dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
+
+if accelerator.is_main_process:
+    print("Total samples", len(dataloader))
+dataloader = accelerator.prepare(dataloader)
+
+start_epoch = 0
+global_step = 0
+total_training_steps = (len(dataloader) * num_epochs)
+world_size = accelerator.state.num_processes
+
+# Загрузка UNet
+latest_checkpoint = os.path.join(checkpoints_folder, project)
+if os.path.isdir(latest_checkpoint):
+    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device, dtype=dtype)
+    if unet_gradient:
+        unet.enable_gradient_checkpointing()
+    unet.set_use_memory_efficient_attention_xformers(False)
+    try:
+        unet.set_attn_processor(AttnProcessor2_0())
+    except Exception as e:
+        print(f"Ошибка при включении SDPA: {e}")
+        unet.set_use_memory_efficient_attention_xformers(True)
+else:
+    raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}")
+
+def create_optimizer(name, params):
+    if name == "adam8bit":
+        return bnb.optim.AdamW8bit(
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01
+        )
+    elif name == "adam":
+        return torch.optim.AdamW(
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01
+        )
+    elif name == "muon_adam8bit":
+        return MuonAdamW8bit(
+            params, 
+            lr=base_learning_rate, 
+            betas=(0.9, betta2), 
+            eps=eps, 
+            weight_decay=0.01,
+            muon_lr_mult=muon_lr_scale,
+        )
+    elif name == "adafactor":
+        return Adafactor(
+            params,
+            lr=base_learning_rate,
+            scale_parameter=False,      # ВАЖНО (иначе авто lr)
+            relative_step=False,        # ВАЖНО
+            warmup_init=False,
+            beta1=0.9,                  # как Adam
+            clip_threshold=1.0,
+            decay_rate=betta2,          # аналог beta2
+            weight_decay=0.01,
+            eps=(1e-30, eps)            # стандарт
+        )
+    else:
+        raise ValueError(f"Unknown optimizer: {name}")
+
+if fbp:
+    trainable_params = list(unet.parameters())
+    optimizer_dict = {p: create_optimizer(optimizer_type, [p]) for p in trainable_params}
+    def optimizer_hook(param):
+        optimizer_dict[param].step()
+        optimizer_dict[param].zero_grad(set_to_none=True)
+    for param in trainable_params:
+        param.register_post_accumulate_grad_hook(optimizer_hook)
+    unet, optimizer = accelerator.prepare(unet, optimizer_dict)
+else:
+    unet.requires_grad_(True)
+    optimizer = create_optimizer(optimizer_type, unet.parameters())
+    
+    def lr_schedule(step):
+        x = step / (total_training_steps * world_size)
+        warmup = warmup_percent
+        if not use_decay:
+            return base_learning_rate
+        if x < warmup:
+            return min_learning_rate + (base_learning_rate - min_learning_rate) * (x / warmup)
+        decay_ratio = (x - warmup) / (1 - warmup)
+        return min_learning_rate + 0.5 * (base_learning_rate - min_learning_rate) * \
+               (1 + math.cos(math.pi * decay_ratio))
+    lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
+    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
+
+if torch_compile:
+    print("compiling")
+    unet = torch.compile(unet)
+    print("compiling - ok")
+
+# Фиксированные семплы (теперь они создаются и лежат на CPU)
+fixed_samples = get_fixed_samples_by_resolution(dataset)
+
+# --- Функция для негативного эмбеддинга ---
+def get_negative_embedding(neg_prompt="", batch_size=1):
+    if not neg_prompt:
+        hidden_dim = 2048 
+        seq_len = 1063 #max_length
+        empty_emb = torch.zeros((batch_size, seq_len, hidden_dim), dtype=torch.float16, device=device)
+        empty_mask = torch.ones((batch_size, seq_len), dtype=torch.bool, device=device)
+        return empty_emb, empty_mask
+
+    uncond_emb, uncond_mask  = encode_texts([neg_prompt])
+    uncond_emb = uncond_emb.to(dtype=dtype, device=device).repeat(batch_size, 1, 1)
+    uncond_mask = uncond_mask.to(device=device).repeat(batch_size, 1)
+
+    return uncond_emb, uncond_mask
+    
+# Получаем негативные (пустые) условия для валидации
+if use_precomputed_embeddings:
+    # 1. грузим encoder ВРЕМЕННО
+    load_text_encoder()
+
+    # 2. считаем negative
+    uncond_emb, uncond_mask = get_negative_embedding()
+
+    # 3. уносим на CPU (очень важно)
+    uncond_emb = uncond_emb.to("cpu")
+    uncond_mask = uncond_mask.to("cpu")
+
+    # 4. выгружаем encoder с GPU
+    del text_encoder
+    torch.cuda.empty_cache()
+    gc.collect()
+
+    text_encoder = None
+
+else:
+    uncond_emb, uncond_mask = get_negative_embedding("low quality")
+    # Уносим негативные эмбеддинги на CPU, чтобы не занимали VRAM во время трейна
+    uncond_emb = uncond_emb.to("cpu")
+    uncond_mask = uncond_mask.to("cpu")
+    
+# --- Функция генерации семплов  ---
+@torch.compiler.disable()
+@torch.no_grad()
+def generate_and_save_samples(fixed_samples_cpu, uncond_data, step):
+    torch.cuda.synchronize()
+    torch.cuda.empty_cache()
+    gc.collect()
+    # 1. Временно переносим негативные эмбеддинги на GPU
+    uncond_emb, uncond_mask = uncond_data
+    uncond_emb = uncond_emb.to(device) 
+    uncond_mask = uncond_mask.to(device)
+    
+    original_model = None
+    try:
+        if not torch_compile:
+            original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
+        else:
+            original_model = unet.eval()
+
+        vae.to(dtype=torch.float16,device=device).eval() 
+        
+        all_generated_images = []
+        all_captions = [] 
+        
+        # 2. Проходим по группам, дергая CPU тензоры и перенося их на GPU только здесь
+        for size, (sample_latents, sample_text_embeddings, sample_mask, sample_text) in fixed_samples_cpu.items():
+            width, height = size
+            # Перенос на GPU происходит в момент вызова .to()
+            sample_latents = sample_latents.to(dtype=dtype, device=device)
+            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
+            sample_mask = sample_mask.to(device=device)
+        
+            latents = torch.randn(
+                sample_latents.shape,
+                device=device,
+                dtype=sample_latents.dtype,
+                generator=torch.Generator(device=device).manual_seed(seed)
+            )
+        
+            scheduler.set_timesteps(n_diffusion_steps, device=device)
+
+            for t in scheduler.timesteps:
+                if guidance_scale != 1:
+                    latent_model_input = torch.cat([latents, latents], dim=0)
+                    
+                    curr_batch_size = sample_text_embeddings.shape[0]
+                    seq_len = sample_text_embeddings.shape[1]
+                    hidden_dim = sample_text_embeddings.shape[2]
+                    
+                    neg_emb_batch = uncond_emb[0:1].expand(curr_batch_size, -1, -1)
+                    text_embeddings_batch = torch.cat([neg_emb_batch, sample_text_embeddings], dim=0)
+                    
+                    neg_mask_batch = uncond_mask[0:1].expand(curr_batch_size, -1)
+                    attention_mask_batch = torch.cat([neg_mask_batch, sample_mask], dim=0)
+                    
+                else:
+                    latent_model_input = latents
+                    text_embeddings_batch = sample_text_embeddings
+                    attention_mask_batch = sample_mask
+            
+                model_out = original_model(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=text_embeddings_batch,
+                    encoder_attention_mask=attention_mask_batch,
+                )
+
+                flow = getattr(model_out, "sample", model_out)
+        
+                if guidance_scale != 1:
+                    flow_uncond, flow_cond = flow.chunk(2)
+                    flow = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
+        
+                latents = scheduler.step(flow, t, latents).prev_sample
+        
+            current_latents = latents
+            if step==0:
+                current_latents = sample_latents
+
+            if latents_mean is not None and latents_std is not None:
+                latents = current_latents * latents_std + latents_mean
+                
+            decoded = vae.decode(latents.to(torch.float16)).sample
+            decoded_fp32 = decoded.to(torch.float32)
+            
+            for img_idx, img_tensor in enumerate(decoded_fp32):
+                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
+                img = img.transpose(1, 2, 0)
+    
+                if np.isnan(img).any():
+                    print("NaNs found, saving stopped! Step:", step)
+                pil_img = Image.fromarray((img * 255).astype("uint8"))
+                
+                max_w_overall = max(s[0] for s in fixed_samples_cpu.keys())
+                max_h_overall = max(s[1] for s in fixed_samples_cpu.keys())
+                max_w_overall = max(255, max_w_overall)
+                max_h_overall = max(255, max_h_overall)
+            
+                padded_img = ImageOps.pad(pil_img, (max_w_overall, max_h_overall), color='white')
+                all_generated_images.append(padded_img)
+
+                caption_text = sample_text[img_idx][:300] if img_idx < len(sample_text) else ""
+                all_captions.append(caption_text)
+                
+                sample_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
+                pil_img.save(sample_path, "JPEG", quality=95)
+        
+        if use_wandb and accelerator.is_main_process:
+            wandb_images = [
+                wandb.Image(img, caption=f"{all_captions[i]}")
+                for i, img in enumerate(all_generated_images)
+            ]
+            wandb.log({"generated_images": wandb_images})
+        if use_comet_ml and accelerator.is_main_process:
+            for i, img in enumerate(all_generated_images):
+                comet_experiment.log_image(
+                    image_data=img,
+                    name=f"step_{step}_img_{i}",
+                    step=step,
+                    metadata={"caption": all_captions[i]}
+                )
+    finally:
+        vae.to("cpu")
+        # Возвращаем негативные эмбеддинги обратно на CPU
+        uncond_emb = uncond_emb.to("cpu")
+        uncond_mask = uncond_mask.to("cpu")
+        try:
+            all_generated_images.clear()
+            all_captions.clear()
+            del all_generated_images, all_captions
+            del latents, current_latents, latent_model_input, flow
+            del decoded, decoded_fp32
+            # Временные копии семплов на GPU удалятся здесь.
+            # Оригиналы в словаре fixed_samples_cpu останутся нетронутыми на CPU.
+            del sample_latents, sample_text_embeddings, sample_mask 
+            del model_out
+        except UnboundLocalError:
+            pass
+            
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# --------------------------- Генерация сэмплов перед обучением ---------------------------
+if accelerator.is_main_process:
+    if save_model:
+        print("Генерация сэмплов до старта обучения...")
+        generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), 0)
+accelerator.wait_for_everyone()
+
+def save_checkpoint(unet, variant=""):
+    if accelerator.is_main_process:
+        model_to_save = None
+        if not torch_compile:
+            model_to_save = accelerator.unwrap_model(unet)
+        else:
+            model_to_save = unet
+
+        if variant != "":
+            model_to_save.to(dtype=torch.float16).save_pretrained(
+                os.path.join(checkpoints_folder, f"{project}"), variant=variant
+            )
+        else:
+            model_to_save.save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
+
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# --------------------------- Тренировочный цикл ---------------------------
+if accelerator.is_main_process:
+    print(f"Total steps per GPU: {total_training_steps}")
+
+epoch_loss_points = []
+progress_bar = tqdm(total=total_training_steps, disable=not accelerator.is_local_main_process, desc="Training", unit="step")
+
+steps_per_epoch = len(dataloader)
+sink_interval = max(1, steps_per_epoch // sink_interval_share)
+min_loss = 4.
+last_sample_time = time.time() 
+sample_interval_seconds = 10 * 60  # 60 минут
+
+for epoch in range(start_epoch, start_epoch + num_epochs):
+    batch_losses = []
+    batch_grads = []
+    batch_sampler.set_epoch(epoch)
+    accelerator.wait_for_everyone()
+    unet.train()
+    
+    for step, (latents, embeddings, attention_mask) in enumerate(dataloader):
+        with accelerator.accumulate(unet):
+            with accelerator.autocast():
+                if save_model == False and epoch == 0 and step == 5 :
+                    used_gb = torch.cuda.max_memory_allocated() / 1024**3
+                    print(f"Шаг {step}: {used_gb:.2f} GB")
+
+                noise = torch.randn_like(latents, dtype=latents.dtype)
+
+                # 3. Время t, bias = -0.5 (Фокус на Деталях ~300) bias = 0.5 (Фокус на структуре) bias = 0 (колокол/ равномерно)
+                bias = 0.1
+                t = torch.sigmoid(torch.randn(latents.shape[0], device=latents.device, dtype=latents.dtype) + bias)
+                
+                # интерполяция между x0 и шумом
+                noisy_latents = (1.0 - t.view(-1, 1, 1, 1)) * latents + t.view(-1, 1, 1, 1) * noise
+                # делаем integer timesteps для UNet
+                timesteps = t.to(torch.float32).mul(999.0)
+                timesteps = timesteps.clamp(0, scheduler.config.num_train_timesteps - 1)
+
+                model_pred = unet(
+                    noisy_latents, 
+                    timesteps, 
+                    encoder_hidden_states=embeddings,
+                    encoder_attention_mask=attention_mask,
+                ).sample
+
+                target = noise - latents
+
+                mse_loss = F.mse_loss(model_pred.float(), target.float())                
+                batch_losses.append(mse_loss.detach().item())
+
+                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
+                    accelerator.wait_for_everyone()
+
+                losses_dict = {}
+                losses_dict["mse"] = mse_loss
+
+                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
+                    accelerator.wait_for_everyone()
+
+                accelerator.backward(mse_loss)
+
+                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
+                    accelerator.wait_for_everyone()
+
+                grad = 0.0
+                if not fbp:
+                    if accelerator.sync_gradients:
+                        grad = 0.
+                        optimizer.step()
+                        lr_scheduler.step()
+                        optimizer.zero_grad(set_to_none=True)
+
+                if accelerator.sync_gradients:
+                    global_step += 1
+                    progress_bar.update(1)
+                    if accelerator.is_main_process:
+                        if fbp:
+                            current_lr = base_learning_rate
+                        else:
+                            current_lr = lr_scheduler.get_last_lr()[0]
+                        batch_grads.append(grad)
+
+                        log_data = {}
+                        log_data["loss_mse"] = mse_loss.detach().item()
+                        log_data["lr"] = current_lr
+                        log_data["grad"] = grad
+                        if accelerator.sync_gradients:
+                            if use_wandb:
+                                wandb.log(log_data, step=global_step)
+                            if use_comet_ml:
+                                comet_experiment.log_metrics(log_data, step=global_step)
+
+                        current_time = time.time()
+                        is_time_to_sample = (current_time - last_sample_time) >= sample_interval_seconds
+                        if is_time_to_sample or global_step == 10:
+                            if save_model:
+                                generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
+                            #elif epoch % 10 == 0:
+                            #    generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
+                            last_n = sink_interval
+
+                            if save_model:
+                                has_losses = len(batch_losses) > 0
+                                avg_sample_loss = np.mean(batch_losses[-sink_interval:]) if has_losses else 0.0
+                                last_loss = batch_losses[-1] if has_losses else 0.0
+                                max_loss = max(avg_sample_loss, last_loss)    
+                                should_save = max_loss < min_loss * save_barrier
+                                print(
+                                    f"Saving: {should_save} | Max: {max_loss:.4f} | "
+                                    f"Last: {last_loss:.4f} | Avg: {avg_sample_loss:.4f}"
+                                )
+                                if should_save:
+                                    min_loss = max_loss
+                                    save_checkpoint(unet)
+                            last_sample_time = current_time
+                            unet.train()
+
+    if accelerator.is_main_process:
+        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses) > 0 else 0.0
+        avg_epoch_grad = np.mean(batch_grads) if len(batch_grads) > 0 else 0.0
+
+        print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
+        log_data_ep = {
+                        "epoch_loss": avg_epoch_loss,
+                        "epoch_grad": avg_epoch_grad,
+                        "epoch": epoch + 1,
+                    }
+        if use_wandb:
+            wandb.log(log_data_ep)
+        if use_comet_ml:
+            comet_experiment.log_metrics(log_data_ep)
+
+if accelerator.is_main_process:
+    print("Обучение завершено! Сохраняем финальную модель...")
+    save_checkpoint(unet,"fp16")
+    if use_comet_ml:
+        comet_experiment.end()
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+    
+print("Готово!")

unet/diffusion_pytorch_model.fp16.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ae783bc49d90ed176212edae361b3ebf33e4f8518a32ca89679a15682f6ac8
+size 3210307232

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:62aa0e71e311b62009ceff8bd7c22f38c6d43a93c792057648b8b5200ab5e014
-size 3210307232
+oid sha256:3e1f8d2466f1a28d6363358e81d3ac4097aaaa5414480d39ae14470cf1faf141
+size 6420443856