muon

requirements.txt

@@ -8,4 +8,5 @@ bitsandbytes>=0.45.4
 transformers
 hf_transfer
 comet_ml
-flash-linear-attention
+flash-linear-attention
+git+https://github.com/recoilme/muon_adamw8bit.git

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2780ff8a377ed48e6fac594ae91ca31d32127771d2662ff5103bdfd42c76fa8e
-size 6429453
+oid sha256:c442f0dceac55de4922225b7cfa939dd38d2e3b4abf2894a45b2e7e098e927fc
+size 6417106

train-Copy1.py

@@ -0,0 +1,1005 @@
+import os
+os.environ["NCCL_P2P_DISABLE"] = "1"
+os.environ["NCCL_IB_DISABLE"] = "1" # comment it on H100
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+import math
+import torch
+import numpy as np
+import matplotlib.pyplot as plt
+from torch.utils.data import DataLoader, Sampler
+from torch.utils.data.distributed import DistributedSampler
+from torch.optim.lr_scheduler import LambdaLR
+from collections import defaultdict
+from diffusers import UNet2DConditionModel,AsymmetricAutoencoderKL,FlowMatchEulerDiscreteScheduler
+from accelerate import Accelerator, DeepSpeedPlugin
+from datasets import load_from_disk
+from tqdm import tqdm
+from PIL import Image, ImageOps
+import wandb
+import random,time
+import gc
+from accelerate.state import DistributedType
+from torch.distributed import broadcast_object_list
+from torch.utils.checkpoint import checkpoint
+from diffusers.models.attention_processor import AttnProcessor2_0
+from datetime import datetime
+import bitsandbytes as bnb
+import torch.nn.functional as F
+from collections import deque
+from transformers import Qwen3_5Tokenizer, Qwen3_5ForConditionalGeneration
+import argparse
+# pip install git+https://github.com/recoilme/muon_adamw8bit.git
+from muon_adamw8bit import MuonAdamW8bit
+
+# --------------------------- Параметры ---------------------------
+ds_path = "datasets/ds1234_noanime_704_vae8x16x"
+project = "unet"
+
+# 1. Считаем локальный батч
+gpu_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
+local_bs = max(1, int((gpu_mem_gb / 32) * 8))
+# 2. Умножаем на количество ГПУ, чтобы получить ГЛОБАЛЬНЫЙ батч
+num_gpus = torch.cuda.device_count()
+batch_size = local_bs * num_gpus
+print(f"GPUs: {num_gpus}, Local BS: {local_bs}, Global BS: {local_bs * num_gpus}")
+base_learning_rate = 1e-5
+min_learning_rate = 1e-7
+num_epochs = num_gpus
+sample_interval_share = 20
+cfg_dropout = 0.10
+max_length = 248
+use_precomputed_embeddings = False
+use_wandb = True
+use_comet_ml = False
+save_model = True
+use_decay = True
+fbp = False
+torch_compile = False
+unet_gradient = True
+loss_normalize = False
+fixed_seed = False
+shuffle = True
+muon_lr_scale = 300
+optimizer_type = "muon_adam8bit"
+if optimizer_type == "muon_adam8bit":
+    batch_size = num_gpus * max(1, int((gpu_mem_gb / 32) * 3))
+comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r" 
+comet_ml_workspace = "recoilme" 
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+# Включение Flash Attention 2/SDPA #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 = 0.01
+betta2 = 0.995
+eps = 1e-7
+clip_grad_norm = 1.0
+limit = 0
+checkpoints_folder = ""
+gradient_accumulation_steps = 1
+dtype = torch.float32
+mixed_precision = "no"
+
+# Параметры для диффузии
+n_diffusion_steps = 40
+samples_to_generate = 12
+guidance_scale = 4
+
+# Папки для сохранения результатов
+generated_folder = "samples"
+os.makedirs(generated_folder, exist_ok=True)
+
+# Настройка seed
+current_date = datetime.now()
+seed = int(current_date.strftime("%Y%m%d")) + 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='Run configuration without saving/sampling')
+
+# 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
+if args.dry_run:
+    save_model = False
+
+# --------------------------- 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(device).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()
+
+def encode_texts(texts, max_length=max_length):
+    load_text_encoder()
+    if texts is None:
+        texts = [""]
+    if isinstance(texts, str):
+        texts = [texts]
+
+    with torch.no_grad():
+        
+        # --- 2. QWEN Энкодер (через Chat Template) ---
+        # 1. Собираем текстовые промпты оборачивая их в Chat Template
+        formatted_prompts = []
+        for t in texts:
+            messages = [{"role": "user", "content": [{"type": "text", "text": t}]}]
+            res_text = tokenizer.apply_chat_template(
+                messages, 
+                add_generation_prompt=True, 
+                tokenize=False 
+            )
+            formatted_prompts.append(res_text)
+
+        # 2. Токенизируем, режем и добавляем паддинг за один раз
+        toks = tokenizer(
+            formatted_prompts,
+            padding="max_length",
+            max_length=max_length,
+            truncation=True,
+            return_tensors="pt"
+        ).to(device)
+
+        # 3. Прогоняем через модель
+        outputs = text_encoder(
+            input_ids=toks.input_ids, 
+            attention_mask=toks.attention_mask,
+            output_hidden_states=True
+        )
+
+        layer_index = -2
+        last_hidden = outputs.hidden_states[layer_index]
+        seq_len = toks.attention_mask.sum(dim=1) - 1
+        pooled = last_hidden[torch.arange(len(last_hidden)), seq_len.clamp(min=0)]
+        # --- НОВАЯ ЛОГИКА: ОБЪЕДИНЕНИЕ ДЛЯ КРОСС-ВНИМАНИЯ ---
+        # 1. Расширяем пулинг-вектор до последовательности [B, 1, 1024]
+        pooled_expanded = pooled.unsqueeze(1) 
+                        
+        # 2. Объединяем последовательность токенов и пулинг-вектор
+        # !!! ИЗМЕНЕНИЕ ЗДЕСЬ !!!: Пулинг идет ПЕРВЫМ
+        # Теперь: [B, 1 + L, 1024]. Пулинг стал токеном в НАЧАЛЕ.
+        new_encoder_hidden_states = torch.cat([pooled_expanded, last_hidden], dim=1) 
+                        
+        # 3. Обновляем маску внимания для нового токена
+        # Маска внимания: [B, 1 + L]. Добавляем 1 в НАЧАЛО.
+        # torch.ones((batch_size, 1), device=device) создает маску [B, 1] со значениями 1.
+        new_attention_mask = torch.cat([torch.ones((last_hidden.shape[0], 1), device=device), toks.attention_mask], dim=1)
+        return new_encoder_hidden_states.to(dtype), new_attention_mask
+        
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+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 ---
+def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
+    size_groups = defaultdict(list)
+    try:
+        widths = dataset["width"]
+        heights = dataset["height"]
+    except KeyError:
+        widths = [0] * len(dataset)
+        heights = [0] * len(dataset)
+    for i, (w, h) in enumerate(zip(widths, heights)):
+        size = (w, h)
+        size_groups[size].append(i)
+    
+    fixed_samples = {}
+    for size, indices in size_groups.items():
+        n_samples = min(samples_per_group, len(indices))
+        if len(size_groups)==1:
+            n_samples = samples_to_generate
+        if n_samples == 0:
+            continue
+        sample_indices = random.sample(indices, n_samples)
+        samples_data = [dataset[idx] for idx in sample_indices]
+        
+        latents = torch.tensor(np.array([item["vae"] for item in samples_data])).to(device=device, dtype=dtype)
+        texts = [item["text"] for item in samples_data]
+        
+        # Кодируем тексты на лету, чтобы получить маски и пулинг
+        #embeddings, masks = encode_texts(texts)
+        if use_precomputed_embeddings:
+            embeddings = torch.tensor(
+                np.array([item["embeddings"] for item in samples_data]),
+                device=device,
+                dtype=dtype
+            )
+            masks = torch.tensor(
+                np.array([item["attention_mask"] for item in samples_data]),
+                device=device,
+                dtype=torch.int64
+            )
+        else:
+            embeddings, masks = encode_texts(texts)
+        
+        fixed_samples[size] = (latents, embeddings, masks, texts)
+    
+    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям")
+    return fixed_samples
+
+if limit > 0:
+    dataset = load_from_disk(ds_path).select(range(limit))
+else:
+    dataset = load_from_disk(ds_path)
+
+
+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=dtype)
+
+    if use_precomputed_embeddings:
+        embeddings = torch.from_numpy(
+            np.array([item["embeddings"] for item in batch], dtype=np.float16)
+        ).to(device, dtype=dtype)
+
+        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.int64)
+
+    return latents, embeddings, attention_mask
+
+batch_sampler = DistributedResolutionBatchSampler(
+        dataset=dataset,
+        batch_size=batch_size,
+        num_replicas=accelerator.num_processes,
+        rank=accelerator.process_index,
+        shuffle = shuffle
+    )
+
+dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
+
+if accelerator.is_main_process:
+    print("Total samples", len(dataloader))
+dataloader = accelerator.prepare(dataloader)
+
+start_epoch = 0
+global_step = 0
+total_training_steps = (len(dataloader) * num_epochs)
+world_size = accelerator.state.num_processes
+
+# Загрузка UNet
+latest_checkpoint = os.path.join(checkpoints_folder, project)
+if os.path.isdir(latest_checkpoint):
+    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device, dtype=dtype)
+    if unet_gradient:
+        unet.enable_gradient_checkpointing()
+    unet.set_use_memory_efficient_attention_xformers(False)
+    try:
+        unet.set_attn_processor(AttnProcessor2_0())
+    except Exception as e:
+        print(f"Ошибка при включении SDPA: {e}")
+        unet.set_use_memory_efficient_attention_xformers(True)
+else:
+    raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}")
+
+# --------------------------- Muon Implementation (FIXED for UNet Conv2d) ---------------------------
+import torch.distributed as dist
+
+def zeropower_via_newtonschulz5(G, steps=10, eps=1e-7):
+    """
+    Оптимизированная версия: поддерживает BF16 и батчинг.
+    """
+    a, b, c = (3.4445, -4.7750, 2.0315)
+    
+    # Используем bfloat16 для скорости (если карта поддерживает, иначе float16/32)
+    # Важно: приведение типа делаем здесь
+    if G.dtype != torch.bfloat16 and G.dtype != torch.float32:
+         # Если пришло float16, лучше считать в float32 или bfloat16 для точности
+         X = G.float() 
+    else:
+         X = G
+
+    # Логика транспонирования (M > N)
+    transposed = G.size(-2) > G.size(-1)
+    if transposed:
+        X = X.mT # .mT - это View, не копирует память
+    
+    # Нормализация по последним двум осям (корректно для любых размерностей)
+    X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps)
+    
+    for _ in range(steps):
+        # A = X @ X.T
+        A = X @ X.mT
+        # B = b*A + c*A@A
+        B = b * A + c * A @ A
+        # X = a*X + B@X
+        X = a * X + B @ X
+        
+    return X.mT if transposed else X
+
+class MuonAdamW8bit(torch.optim.Optimizer):
+    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01, 
+                 muon_lr=None, muon_momentum=0.95, ns_steps=5):
+        params = list(params)
+        
+        matrix_params = []
+        scalar_params = []
+        
+        for p in params:
+            if p.ndim >= 2:
+                matrix_params.append(p)
+            else:
+                scalar_params.append(p)
+        
+        # --- ВАЖНОЕ ИСПРАВЛЕНИЕ LR ---
+        # Если базовый lr=1e-5, Muon должен быть ~0.01 -> 0.02
+        # Ставим коэф x1000 или x2000 от базового, или задаем явно.
+        # Автор ставит 0.02. Давайте поставим 0.01 как безопасный минимум для старта.
+        actual_muon_lr = muon_lr if muon_lr is not None else lr * muon_lr_scale
+        self.muon_lr_scale = actual_muon_lr / lr if lr > 0 else 1.0
+        self.ns_steps = ns_steps
+        
+        # Инициализируем Muon (передаем ему только матрицы)
+        # Обратите внимание: lr передаем актуальный
+        self.muon_opt = MuonInternal(matrix_params, lr=actual_muon_lr, momentum=muon_momentum, ns_steps=ns_steps)
+        
+        # Adam для скаляров (biases, norms)
+        self.adam_opt = bnb.optim.AdamW8bit(scalar_params, lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
+
+        self.defaults = dict(lr=lr) 
+        self.state = {}
+        self.param_groups = self.adam_opt.param_groups
+
+    def zero_grad(self, set_to_none=False):
+        self.muon_opt.zero_grad(set_to_none)
+        self.adam_opt.zero_grad(set_to_none)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        # Синхронизация LR (если шедулер меняет lr в adam группах)
+        current_base_lr = self.adam_opt.param_groups[0]['lr']
+        for group in self.muon_opt.param_groups:
+            group['lr'] = current_base_lr * self.muon_lr_scale
+            
+        self.muon_opt.step(closure)
+        self.adam_opt.step(closure)
+
+    # state_dict/load_state_dict оставляем как было у вас
+    def state_dict(self):
+        return {
+            'muon': self.muon_opt.state_dict(),
+            'adam': self.adam_opt.state_dict()
+        }
+
+    def load_state_dict(self, state_dict):
+        self.muon_opt.load_state_dict(state_dict['muon'])
+        self.adam_opt.load_state_dict(state_dict['adam'])
+
+
+# Внутренний класс Muon (логика обновления матриц)
+class MuonInternal(torch.optim.Optimizer):
+    def __init__(self, params, lr=0.01, momentum=0.95, ns_steps=5):
+        super().__init__(params, dict(lr=lr, momentum=momentum, ns_steps=ns_steps))
+        self.distributed = torch.distributed.is_initialized()
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            lr = group["lr"]
+            momentum = group["momentum"]
+            ns_steps = group["ns_steps"]
+
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+                
+                g = p.grad
+                state = self.state[p]
+                
+                if "momentum_buffer" not in state:
+                    state["momentum_buffer"] = torch.zeros_like(g, dtype=torch.float32)
+                
+                buf = state["momentum_buffer"]
+                
+                # Обновление момента
+                buf.mul_(momentum).add_(g.float())
+                
+                # Nesterov
+                update = g.float().add(buf, alpha=momentum)
+                
+                # Обработка Conv2d (flatten)
+                original_shape = g.shape
+                if g.ndim == 4:
+                    update = update.view(update.size(0), -1)
+                
+                # Ортогонализация (теперь принимает update и steps)
+                update = zeropower_via_newtonschulz5(update, steps=ns_steps)
+                
+                # Масштабирование (стандарт для Muon)
+                update *= max(1, update.size(-2) / update.size(-1)) ** 0.5
+                
+                # Возврат формы для Conv
+                if g.ndim == 4:
+                    update = update.view(original_shape)
+                    
+                p.add_(update, alpha=-lr)
+        return loss
+
+# --------------------------- End Muon Implementation ---------------------------
+
+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_momentum=0.95,
+            ns_dtype=torch.bfloat16,
+            muon_lr_mult=1000.0,
+        )
+    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())
+    # 1. Сначала замораживаем ВСЕ параметры UNet
+    #unet.requires_grad_(False)
+    
+    # 2. Размораживаем только нужные
+    #trainable_params_names = ["conv_in.weight", "conv_in.bias", "conv_out.weight", "conv_out.bias"]
+    #train_params = []
+    
+    #for name, param in unet.named_parameters():
+    #    if any(target in name for target in trainable_params_names):
+    #        param.requires_grad = True
+    #        train_params.append(param)
+    #        print(f"Обучаемый слой: {name}")
+    
+    def lr_schedule(step):
+        x = step / (total_training_steps * world_size)
+        warmup = warmup_percent
+        if not use_decay:
+            return base_learning_rate
+        if x < warmup:
+            return min_learning_rate + (base_learning_rate - min_learning_rate) * (x / warmup)
+        decay_ratio = (x - warmup) / (1 - warmup)
+        return min_learning_rate + 0.5 * (base_learning_rate - min_learning_rate) * \
+               (1 + math.cos(math.pi * decay_ratio))
+    lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
+    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
+
+if torch_compile:
+    print("compiling")
+    unet = torch.compile(unet)
+    print("compiling - ok")
+
+# Фиксированные семплы
+fixed_samples = get_fixed_samples_by_resolution(dataset)
+
+# --- [UPDATED] Функция для негативного эмбеддинга (возвращает 3 элемента) ---
+def get_negative_embedding(neg_prompt="", batch_size=1):
+    if not neg_prompt:
+        hidden_dim = 2048 
+        seq_len = max_length
+        empty_emb = torch.zeros((batch_size, seq_len, hidden_dim), dtype=dtype, device=device)
+        empty_mask = torch.ones((batch_size, seq_len), dtype=torch.int64, device=device)
+        return empty_emb, empty_mask
+
+    uncond_emb, uncond_mask  = encode_texts([neg_prompt])
+    uncond_emb = uncond_emb.to(dtype=dtype, device=device).repeat(batch_size, 1, 1)
+    uncond_mask = uncond_mask.to(device=device).repeat(batch_size, 1)
+
+    return uncond_emb, uncond_mask
+    
+# Получаем негативные (пустые) условия для валидации
+if use_precomputed_embeddings:
+    # 1. грузим encoder ВРЕМЕННО
+    load_text_encoder()
+
+    # 2. считаем negative
+    uncond_emb, uncond_mask = get_negative_embedding("low quality")
+
+    # 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")
+    
+# --- Функция генерации семплов  ---
+@torch.compiler.disable()
+@torch.no_grad()
+def generate_and_save_samples(fixed_samples_cpu, uncond_data, step):
+    uncond_emb, uncond_mask = uncond_data
+    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(device=device).eval() 
+        
+        all_generated_images = []
+        all_captions = [] 
+        
+        # Распаковываем 5 элементов (добавились mask)
+        for size, (sample_latents, sample_text_embeddings, sample_mask, sample_text) in fixed_samples_cpu.items():
+            width, height = size
+            sample_latents = sample_latents.to(dtype=dtype, device=device)
+            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
+            sample_mask = sample_mask.to(device=device)
+        
+            latents = torch.randn(
+                sample_latents.shape,
+                device=device,
+                dtype=sample_latents.dtype,
+                generator=torch.Generator(device=device).manual_seed(seed)
+            )
+        
+            scheduler.set_timesteps(n_diffusion_steps, device=device)
+
+            for t in scheduler.timesteps:
+                if guidance_scale != 1:
+                    latent_model_input = torch.cat([latents, latents], dim=0)
+                    
+                    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
+            
+                # Теперь всё имеет одинаковый batch size
+                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.float32)).sample
+            decoded_fp32 = decoded.to(torch.float32)
+            
+            for img_idx, img_tensor in enumerate(decoded_fp32):
+                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
+                img = img.transpose(1, 2, 0)
+    
+                if np.isnan(img).any():
+                    print("NaNs found, saving stopped! Step:", step)
+                pil_img = Image.fromarray((img * 255).astype("uint8"))
+                
+                max_w_overall = max(s[0] for s in fixed_samples_cpu.keys())
+                max_h_overall = max(s[1] for s in fixed_samples_cpu.keys())
+                max_w_overall = max(255, max_w_overall)
+                max_h_overall = max(255, max_h_overall)
+            
+                padded_img = ImageOps.pad(pil_img, (max_w_overall, max_h_overall), color='white')
+                all_generated_images.append(padded_img)
+
+                caption_text = sample_text[img_idx][:300] if img_idx < len(sample_text) else ""
+                all_captions.append(caption_text)
+                
+                sample_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
+                pil_img.save(sample_path, "JPEG", quality=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")
+        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
+            del sample_latents, sample_text_embeddings, sample_mask # Копии на GPU
+            del model_out
+        except UnboundLocalError:
+            pass
+            
+        # 3. Синхронизируем CUDA перед очисткой
+        torch.cuda.synchronize()
+        # 4. Теперь чистим кэш аллокатора и вызываем GC
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# --------------------------- 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)
+sample_interval = max(1, steps_per_epoch // sample_interval_share)
+min_loss = 4.
+last_sample_time = time.time() 
+sample_interval_seconds = 60 * 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):
+            if save_model == False and epoch == 0 and step == 5 :
+                used_gb = torch.cuda.max_memory_allocated() / 1024**3
+                print(f"Шаг {step}: {used_gb:.2f} GB")
+                
+            # шум
+            noise = torch.randn_like(latents, dtype=latents.dtype)
+
+            # 3. Время t (сэмплим, как и раньше, но чуть сжимаем края)
+            u = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            t = u * (1 - 2 * 1e-5) + 1e-5  # Теперь t строго в (0.00001 ... 0.99999)
+            # интерполяция между x0 и шумом
+            noisy_latents = (1.0 - t.view(-1, 1, 1, 1)) * latents + t.view(-1, 1, 1, 1) * noise
+            # делаем integer timesteps для UNet
+            timesteps = t.to(torch.float32).mul(999.0)
+            timesteps = timesteps.clamp(0, scheduler.config.num_train_timesteps - 1)
+
+            # --- Вызов UNet с маской  ---
+            model_pred = unet(
+                noisy_latents, 
+                timesteps, 
+                encoder_hidden_states=embeddings,
+                encoder_attention_mask=attention_mask,
+            ).sample
+            
+            target = noise - latents
+
+            mse_loss = F.mse_loss(model_pred.float(), target.float())                
+            batch_losses.append(mse_loss.detach().item())
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+
+            losses_dict = {}
+            losses_dict["mse"] = mse_loss
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+
+            accelerator.backward(mse_loss)
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+                
+            grad = 0.0
+            if not fbp:
+                if accelerator.sync_gradients:
+                    grad_val = accelerator.clip_grad_norm_(unet.parameters(), clip_grad_norm)
+                    grad = grad_val.float().item() if torch.is_tensor(grad_val) else float(grad_val)
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+
+            if accelerator.sync_gradients:
+                global_step += 1
+                progress_bar.update(1)
+                if accelerator.is_main_process:
+                    if fbp:
+                        current_lr = base_learning_rate
+                    else:
+                        current_lr = lr_scheduler.get_last_lr()[0]
+                    batch_grads.append(grad)
+    
+                    log_data = {}
+                    log_data["loss_mse"] = mse_loss.detach().item()
+                    log_data["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 == 50:
+                        # Передаем tuple (emb, mask) для негатива
+                        if save_model:
+                            generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
+                        elif epoch % 10 == 0:
+                            generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
+                        last_n = sample_interval
+                        
+                        if save_model:
+                            has_losses = len(batch_losses) > 0
+                            avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if has_losses else 0.0
+                            last_loss = batch_losses[-1] if has_losses else 0.0
+                            max_loss = max(avg_sample_loss, last_loss)    
+                            should_save = max_loss < min_loss * save_barrier
+                            print(
+                                f"Saving: {should_save} | Max: {max_loss:.4f} | "
+                                f"Last: {last_loss:.4f} | Avg: {avg_sample_loss:.4f}"
+                            )
+                            # 6. Сохранение и обновление
+                            if should_save:
+                                min_loss = max_loss
+                                save_checkpoint(unet)
+                        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...")
+    #if save_model:
+    save_checkpoint(unet,"fp16")
+    if use_comet_ml:
+        comet_experiment.end()
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+    
+print("Готово!")

train.py

@@ -28,6 +28,8 @@ import torch.nn.functional as F
 from collections import deque
 from transformers import Qwen3_5Tokenizer, Qwen3_5ForConditionalGeneration
 import argparse
+# pip install git+https://github.com/recoilme/muon_adamw8bit.git
+from muon_adamw8bit import MuonAdamW8bit
 
 # --------------------------- Параметры ---------------------------
 ds_path = "datasets/ds1234_noanime_704_vae8x16x"
@@ -40,8 +42,8 @@ local_bs = max(1, int((gpu_mem_gb / 32) * 8))
 num_gpus = torch.cuda.device_count()
 batch_size = local_bs * num_gpus
 print(f"GPUs: {num_gpus}, Local BS: {local_bs}, Global BS: {local_bs * num_gpus}")
-base_learning_rate = 1e-5
-min_learning_rate = 1e-7
+base_learning_rate = 4e-5
+min_learning_rate = 4e-7
 num_epochs = num_gpus
 sample_interval_share = 20
 cfg_dropout = 0.10
@@ -60,7 +62,7 @@ shuffle = True
 muon_lr_scale = 300
 optimizer_type = "muon_adam8bit"
 if optimizer_type == "muon_adam8bit":
-    batch_size = num_gpus * max(1, int((gpu_mem_gb / 32) * 3))
+    batch_size = num_gpus * max(1, int((gpu_mem_gb / 32) * 4))
 comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r" 
 comet_ml_workspace = "recoilme" 
 torch.backends.cuda.matmul.allow_tf32 = True
@@ -434,156 +436,6 @@ if os.path.isdir(latest_checkpoint):
 else:
     raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}")
 
-# --------------------------- Muon Implementation (FIXED for UNet Conv2d) ---------------------------
-import torch.distributed as dist
-
-def zeropower_via_newtonschulz5(G, steps=10, eps=1e-7):
-    """
-    Оптимизированная версия: поддерживает BF16 и батчинг.
-    """
-    a, b, c = (3.4445, -4.7750, 2.0315)
-    
-    # Используем bfloat16 для скорости (если карта поддерживает, иначе float16/32)
-    # Важно: приведение типа делаем здесь
-    if G.dtype != torch.bfloat16 and G.dtype != torch.float32:
-         # Если пришло float16, лучше считать в float32 или bfloat16 для точности
-         X = G.float() 
-    else:
-         X = G
-
-    # Логика транспонирования (M > N)
-    transposed = G.size(-2) > G.size(-1)
-    if transposed:
-        X = X.mT # .mT - это View, не копирует память
-    
-    # Нормализация по последним двум осям (корректно для любых размерностей)
-    X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps)
-    
-    for _ in range(steps):
-        # A = X @ X.T
-        A = X @ X.mT
-        # B = b*A + c*A@A
-        B = b * A + c * A @ A
-        # X = a*X + B@X
-        X = a * X + B @ X
-        
-    return X.mT if transposed else X
-
-class MuonAdamW8bit(torch.optim.Optimizer):
-    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01, 
-                 muon_lr=None, muon_momentum=0.95, ns_steps=5):
-        params = list(params)
-        
-        matrix_params = []
-        scalar_params = []
-        
-        for p in params:
-            if p.ndim >= 2:
-                matrix_params.append(p)
-            else:
-                scalar_params.append(p)
-        
-        # --- ВАЖНОЕ ИСПРАВЛЕНИЕ LR ---
-        # Если базовый lr=1e-5, Muon должен быть ~0.01 -> 0.02
-        # Ставим коэф x1000 или x2000 от базового, или задаем явно.
-        # Автор ставит 0.02. Давайте поставим 0.01 как безопасный минимум для старта.
-        actual_muon_lr = muon_lr if muon_lr is not None else lr * muon_lr_scale
-        self.muon_lr_scale = actual_muon_lr / lr if lr > 0 else 1.0
-        self.ns_steps = ns_steps
-        
-        # Инициализируем Muon (передаем ему только матрицы)
-        # Обратите внимание: lr передаем актуальный
-        self.muon_opt = MuonInternal(matrix_params, lr=actual_muon_lr, momentum=muon_momentum, ns_steps=ns_steps)
-        
-        # Adam для скаляров (biases, norms)
-        self.adam_opt = bnb.optim.AdamW8bit(scalar_params, lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
-
-        self.defaults = dict(lr=lr) 
-        self.state = {}
-        self.param_groups = self.adam_opt.param_groups
-
-    def zero_grad(self, set_to_none=False):
-        self.muon_opt.zero_grad(set_to_none)
-        self.adam_opt.zero_grad(set_to_none)
-
-    @torch.no_grad()
-    def step(self, closure=None):
-        # Синхронизация LR (если шедулер меняет lr в adam группах)
-        current_base_lr = self.adam_opt.param_groups[0]['lr']
-        for group in self.muon_opt.param_groups:
-            group['lr'] = current_base_lr * self.muon_lr_scale
-            
-        self.muon_opt.step(closure)
-        self.adam_opt.step(closure)
-
-    # state_dict/load_state_dict оставляем как было у вас
-    def state_dict(self):
-        return {
-            'muon': self.muon_opt.state_dict(),
-            'adam': self.adam_opt.state_dict()
-        }
-
-    def load_state_dict(self, state_dict):
-        self.muon_opt.load_state_dict(state_dict['muon'])
-        self.adam_opt.load_state_dict(state_dict['adam'])
-
-
-# Внутренний класс Muon (логика обновления матриц)
-class MuonInternal(torch.optim.Optimizer):
-    def __init__(self, params, lr=0.01, momentum=0.95, ns_steps=5):
-        super().__init__(params, dict(lr=lr, momentum=momentum, ns_steps=ns_steps))
-        self.distributed = torch.distributed.is_initialized()
-
-    @torch.no_grad()
-    def step(self, closure=None):
-        loss = None
-        if closure is not None:
-            with torch.enable_grad():
-                loss = closure()
-
-        for group in self.param_groups:
-            lr = group["lr"]
-            momentum = group["momentum"]
-            ns_steps = group["ns_steps"]
-
-            for p in group["params"]:
-                if p.grad is None:
-                    continue
-                
-                g = p.grad
-                state = self.state[p]
-                
-                if "momentum_buffer" not in state:
-                    state["momentum_buffer"] = torch.zeros_like(g, dtype=torch.float32)
-                
-                buf = state["momentum_buffer"]
-                
-                # Обновление момента
-                buf.mul_(momentum).add_(g.float())
-                
-                # Nesterov
-                update = g.float().add(buf, alpha=momentum)
-                
-                # Обработка Conv2d (flatten)
-                original_shape = g.shape
-                if g.ndim == 4:
-                    update = update.view(update.size(0), -1)
-                
-                # Ортогонализация (теперь принимает update и steps)
-                update = zeropower_via_newtonschulz5(update, steps=ns_steps)
-                
-                # Масштабирование (стандарт для Muon)
-                update *= max(1, update.size(-2) / update.size(-1)) ** 0.5
-                
-                # Возврат формы для Conv
-                if g.ndim == 4:
-                    update = update.view(original_shape)
-                    
-                p.add_(update, alpha=-lr)
-        return loss
-
-# --------------------------- End Muon Implementation ---------------------------
-
 def create_optimizer(name, params):
     if name == "adam8bit":
         return bnb.optim.AdamW8bit(
@@ -600,7 +452,9 @@ def create_optimizer(name, params):
             betas=(0.9, betta2), 
             eps=eps, 
             weight_decay=0.01,
-            muon_momentum=0.95
+            muon_momentum=0.95,
+            ns_dtype=torch.bfloat16,
+            muon_lr_mult=1000.0,
         )
     else:
         raise ValueError(f"Unknown optimizer: {name}")

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:24a108643923cb0dc32d1be767a8004df4a4cc5dfd880464b864bd2b4b841fc9
+oid sha256:feb68ae24f6d1ada2f50a1bbd849b45f5ffd116c6bc0438b28ef9f1bd4cbd3cd
 size 6294042336