undo vae unet

src/model_create.ipynb

@@ -603,7 +603,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "1cb4ff0f-36cc-43cf-86a4-aaab9f106725",
+   "id": "0a1d32f6-23ea-4f6f-b2e8-9584a0c12a0d",
    "metadata": {},
    "outputs": [],
    "source": []

train.py

@@ -7,11 +7,12 @@ from torch.utils.data import DataLoader, Sampler
 from torch.utils.data.distributed import DistributedSampler
 from torch.optim.lr_scheduler import LambdaLR
 from collections import defaultdict
-from diffusers import UNet2DConditionModel, AutoencoderKL
+from torch.optim.lr_scheduler import LambdaLR
+from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL
 from accelerate import Accelerator
 from datasets import load_from_disk
 from tqdm import tqdm
-from PIL import Image, ImageOps
+from PIL import Image,ImageOps
 import wandb
 import random
 import gc
@@ -30,7 +31,7 @@ project = "unet"
 batch_size = 128
 base_learning_rate = 9e-5
 min_learning_rate = 1e-5
-num_epochs = 84
+num_epochs = 63
 # samples/save per epoch
 sample_interval_share = 5
 use_wandb = True
@@ -44,8 +45,8 @@ unet_gradient = True
 clip_sample = False #Scheduler
 fixed_seed = True
 shuffle = True
-comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"
-comet_ml_workspace = "recoilme"
+comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"  # Добавлен API ключ для Comet ML
+comet_ml_workspace = "recoilme"  # Добавлен workspace для Comet ML
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
@@ -59,7 +60,7 @@ clip_grad_norm = 1.0
 steps_offset = 0 # Scheduler
 limit = 0
 checkpoints_folder = ""
-mixed_precision = "no" # "fp16"
+mixed_precision = "no" #"fp16"
 gradient_accumulation_steps = 1
 accelerator = Accelerator(
     mixed_precision=mixed_precision,
@@ -87,7 +88,7 @@ if fixed_seed:
         torch.cuda.manual_seed_all(seed)
 
 # --------------------------- Параметры LoRA ---------------------------
-lora_name = ""
+lora_name = "" 
 lora_rank = 32
 lora_alpha = 64
 
@@ -102,6 +103,13 @@ def sample_timesteps_bias(
     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":
@@ -118,16 +126,17 @@ def sample_timesteps_bias(
     timesteps = steps_offset + (samples * max_idx).long().to(device)
     return timesteps
 
-
 def logit_normal_samples(shape, mu=0.0, sigma=1.0, device=None, dtype=None):
     normal_samples = torch.normal(mean=mu, std=sigma, size=shape, device=device, dtype=dtype)
+
     logit_normal_samples = torch.sigmoid(normal_samples)
+
     return logit_normal_samples
 
 # --------------------------- Инициализация WandB ---------------------------
 if accelerator.is_main_process:
     if use_wandb:
-        wandb.init(project=project + lora_name, config={
+        wandb.init(project=project+lora_name, config={
             "batch_size": batch_size,
             "base_learning_rate": base_learning_rate,
             "num_epochs": num_epochs,
@@ -141,6 +150,7 @@ if accelerator.is_main_process:
             project_name=project,
             workspace=comet_ml_workspace
         )
+        # Логируем гиперпараметры в Comet ML
         hyper_params = {
             "batch_size": batch_size,
             "base_learning_rate": base_learning_rate,
@@ -160,6 +170,7 @@ 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 = AutoencoderKL.from_pretrained("AiArtLab/simplevae", subfolder="vae", torch_dtype=dtype).to("cpu").eval()
 
 shift_factor = getattr(vae.config, "shift_factor", 0.0)
@@ -174,6 +185,8 @@ latents_mean = getattr(vae.config, "latents_mean", None)
 latents_std = getattr(vae.config, "latents_std", None)
 
 
+
+
 class DistributedResolutionBatchSampler(Sampler):
     def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
         self.dataset = dataset
@@ -253,16 +266,14 @@ def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
     fixed_samples = {}
     for size, indices in size_groups.items():
         n_samples = min(samples_per_group, len(indices))
-        if len(size_groups) == 1:
+        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]
-
-        # FIXED: keep fixed samples on CPU to avoid device/device-transfer issues when creating dataset
-        latents = torch.tensor(np.array([item["vae"] for item in samples_data]), dtype=dtype).cpu()
-        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data]), dtype=dtype).cpu()
+        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)
     
@@ -274,10 +285,9 @@ if limit > 0:
 else:
     dataset = load_from_disk(ds_path)
 
-
 def collate_fn_simple(batch):
-    latents = torch.tensor(np.array([item["vae"] for item in batch]), dtype=dtype).to(device)
-    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch]), dtype=dtype).to(device)
+    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(
@@ -288,23 +298,20 @@ batch_sampler = DistributedResolutionBatchSampler(
         shuffle=shuffle
     )
 
-# NOTE: we create dataloader first, then prepare it with accelerator. We'll create optimizer/lr_scheduler after
-# we know len(dataloader) (which is per-process after prepare) so that scheduling is consistent.
-
 dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
-print("Total batches (pre-prepare):", len(dataloader))
+print("Total samples",len(dataloader))
 dataloader = accelerator.prepare(dataloader)
 
-# --------------------------- Теперь безопасно --- вычисляем шаги и создаём оптимизатор/шедулер ---------------------------
-steps_per_epoch = len(dataloader)  # this is per-process (after prepare)
-total_training_steps = steps_per_epoch * num_epochs
-print(f"[rank {accelerator.process_index}] steps_per_epoch={steps_per_epoch}, total_training_steps={total_training_steps}")
+start_epoch = 0
+global_step = 0
+total_training_steps = (len(dataloader) * num_epochs)
+world_size = accelerator.state.num_processes
 
-# --------------------------- Создание/загрузка UNet ---------------------------
+# 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)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device,dtype=dtype)
     if torch_compile:
         print("compiling")
         torch.set_float32_matmul_precision('high')
@@ -318,14 +325,16 @@ if os.path.isdir(latest_checkpoint):
     except Exception as e:
         print(f"Ошибка при включении SDPA: {e}")
         unet.set_use_memory_efficient_attention_xformers(True)
+
 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.")
 
-# --------------------------- LoRA (если нужно) ---------------------------
 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.")
 
@@ -363,8 +372,8 @@ if lora_name:
 if lora_name:
     trainable_params = [p for p in unet.parameters() if p.requires_grad]
 else:
-    trainable_params = list(unet.parameters()) if fbp else [p for p in unet.parameters() if p.requires_grad]
-
+    if fbp:
+        trainable_params = list(unet.parameters())
 
 def create_optimizer(name, params):
     if name == "adam8bit":
@@ -376,18 +385,6 @@ def create_optimizer(name, params):
         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}")
 
@@ -398,15 +395,11 @@ if fbp:
         optimizer_dict[param].zero_grad(set_to_none=True)
     for param in trainable_params:
         param.register_post_accumulate_grad_hook(optimizer_hook)
-    # FIXED: prepare fbp variant (keeps original logic)
     unet, optimizer = accelerator.prepare(unet, optimizer_dict)
 else:
-    optimizer = create_optimizer(optimizer_type, trainable_params)
-
-    # FIXED: LR schedule should be based on total_training_steps (per-process steps * epochs)
+    optimizer = create_optimizer(optimizer_type, unet.parameters())
     def lr_schedule(step):
-        # step is current scheduler step (0..total_training_steps)
-        x = step / max(1, total_training_steps)
+        x = step / (total_training_steps * world_size)
         warmup = warmup_percent
         if not use_decay:
             return base_learning_rate
@@ -415,8 +408,6 @@ else:
         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))
-
-    # LambdaLR expects a multiplier, so divide by base_learning_rate
     lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
 
     num_params = sum(p.numel() for p in unet.parameters())
@@ -424,13 +415,12 @@ else:
     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}")
-
-    # FIXED: prepare model, optimizer, scheduler AFTER creating them and after dataloader.prepare
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
 
 # --------------------------- 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
@@ -438,12 +428,12 @@ def generate_and_save_samples(fixed_samples_cpu, step):
         original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
         vae.to(device=device).eval()  # временно подгружаем VAE на GPU для декодинга
 
+        
         all_generated_images = []
         all_captions = [] 
         
         for size, (sample_latents, sample_text_embeddings, sample_text) in fixed_samples_cpu.items():
             width, height = size
-            # move CPU tensors to device here (they were kept on CPU in get_fixed_samples_by_resolution)
             sample_latents = sample_latents.to(dtype=dtype, device=device)
             sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
 
@@ -482,11 +472,19 @@ def generate_and_save_samples(fixed_samples_cpu, step):
 
             # Параметры нормализации
             latent_for_vae = current_latents.detach() / scaling_factor + shift_factor
+            
             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)
+                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"))
@@ -510,7 +508,7 @@ def generate_and_save_samples(fixed_samples_cpu, step):
                 wandb.Image(img, caption=f"{all_captions[i]}")
                 for i, img in enumerate(all_generated_images)
             ]
-            wandb.log({"generated_images": wandb_images}, step=step)
+            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(
@@ -525,7 +523,11 @@ def generate_and_save_samples(fixed_samples_cpu, 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()
 
@@ -533,23 +535,20 @@ def generate_and_save_samples(fixed_samples_cpu, step):
 if accelerator.is_main_process:
     if save_model:
         print("Генерация сэмплов до старта обучения...")
-        generate_and_save_samples(fixed_samples, 0)
+        generate_and_save_samples(fixed_samples,0)
 accelerator.wait_for_everyone()
 
 # Модиф��цируем функцию сохранения модели для поддержки LoRA
-def save_checkpoint(unet, variant=""):
+def save_checkpoint(unet,variant=""):
     if accelerator.is_main_process:
         if lora_name:
             save_lora_checkpoint(unet)
         else:
-            # FIXED: don't change dtype of model wrapped by accelerator. Unwrap and save as-is.
-            model_to_save = accelerator.unwrap_model(unet)
-            dest = os.path.join(checkpoints_folder, f"{project}")
-            if variant != "":
-                model_to_save.save_pretrained(dest, variant=variant)
+            if variant!="":
+                accelerator.unwrap_model(unet.to(dtype=torch.float16)).save_pretrained(os.path.join(checkpoints_folder, f"{project}"),variant=variant)
             else:
-                model_to_save.save_pretrained(dest)
-
+                accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
+            unet = unet.to(dtype=dtype)
 
 # --------------------------- Тренировочный цикл ---------------------------
 if accelerator.is_main_process:
@@ -558,25 +557,20 @@ if accelerator.is_main_process:
 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 = 2.
 
-for epoch in range(0, num_epochs):
-    # FIXED: set epoch on the dataloader's batch_sampler if available (accelerator may have wrapped it)
-    if hasattr(dataloader, "batch_sampler") and hasattr(dataloader.batch_sampler, "set_epoch"):
-        dataloader.batch_sampler.set_epoch(epoch)
-    elif hasattr(batch_sampler, "set_epoch"):
-        batch_sampler.set_epoch(epoch)
-
-    accelerator.wait_for_everyone()
-    unet.train()
-
+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()
+    #print("epoch:",epoch)
     for step, (latents, embeddings) in enumerate(dataloader):
         with accelerator.accumulate(unet):
-            if save_model == False and step == 5:
+            if save_model == False and step == 5 :
                 used_gb = torch.cuda.max_memory_allocated() / 1024**3
                 print(f"Шаг {step}: {used_gb:.2f} GB")
                 
@@ -590,94 +584,83 @@ for epoch in range(0, num_epochs):
 
             mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
 
-            # Сохраняем для локальных логов
-            batch_losses.append(mse_loss.detach().cpu().item())
+            # 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(mse_loss)
 
-            grad_norm_val = 0.0
+            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:
-                    # Clip gradients and step only when gradients are synchronized (i.e. actual optimizer step)
-                    grad_norm = accelerator.clip_grad_norm_(unet.parameters(), clip_grad_norm)
-                    grad_norm_val = float(grad_norm)
+                    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)
 
-            # increment global_step only when we have synchronized gradients (i.e. on optimizer step)
-            # FIXED: ensure global_step reflects optimizer updates, not micro-batches
             if accelerator.sync_gradients:
-                try:
-                    global_step += 1
-                except NameError:
-                    global_step = 1
+                global_step += 1
                 progress_bar.update(1)
-
-                # Aggregate loss across processes for correct logging
-                loss_tensor = mse_loss.detach().clone()
-                # move to device if not already
-                loss_tensor = loss_tensor.to(device)
-                gathered = accelerator.gather(loss_tensor)
-                if accelerator.is_main_process:
-                    reduced_loss = gathered.mean().item()
-                else:
-                    reduced_loss = None
-
-                # Логируем метрики только на главном процессе
+                # Логируем метрики
                 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_norm_val)
-
-                    log_data = {
-                        "loss": reduced_loss,
-                        "lr": current_lr,
-                        "grad": grad_norm_val,
-                        "epoch": epoch + 1,
-                        "global_step": global_step,
-                    }
-
-                    if use_wandb:
-                        wandb.log(log_data, step=global_step)
-                    if use_comet_ml:
-                        comet_experiment.log_metrics(log_data, step=global_step)
-
-                    # Генерируем сэмплы с заданным интервалом (только на главном процессе)
+                    batch_grads.append(grad)
+    
+                    log_data = {}
+                    log_data["loss"] = 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)
+    
+                    # Генерируем сэмплы с заданным интервалом
                     if global_step % sample_interval == 0:
-                        generate_and_save_samples(fixed_samples, global_step)
-
+                        generate_and_save_samples(fixed_samples,global_step)
+                        last_n = sample_interval
+    
                         if save_model:
-                            # use recent local losses to decide saving (still local); you may want to use reduced_loss here
                             avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if len(batch_losses) > 0 else 0.0
-                            if use_wandb:
-                                wandb.log({"sample_loss": avg_sample_loss)
-                            print("saving:", reduced_loss is not None and reduced_loss < min_loss * save_barrier, "Current:", reduced_loss, "Avg:", avg_sample_loss)
-                            if reduced_loss is not None and reduced_loss < min_loss * save_barrier:
+                            print("saving:", avg_sample_loss < min_loss * save_barrier, "Avg:", avg_sample_loss)
+                            if avg_sample_loss is not None and avg_sample_loss < min_loss * save_barrier:
                                 min_loss = avg_sample_loss
                                 save_checkpoint(unet)
 
-    # Эпоха окончена — агрегируем и логируем средние значения
+
     if accelerator.is_main_process:
         # local averages
         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
 
-        # FIXED: optionally reduce across processes if you want a true global epoch average
-        # Here we compute local values and log them (main process only). For global average use accelerator.gather
-        print(f"\nЭпоха {epoch} завершена. Средний лосс (local main proc): {avg_epoch_loss:.6f}")
+        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({"epoch_loss": avg_epoch_loss, "epoch_grad": avg_epoch_grad, "epoch": epoch + 1})
+            wandb.log(log_data_ep)
+        if use_comet_ml:
+            comet_experiment.log_metrics(log_data_ep)
 
 # Завершение обучения - сохраняем финальную модель
 if accelerator.is_main_process:
     print("Обучение завершено! Сохраняем финальную модель...")
     if save_model:
-        save_checkpoint(unet, "fp16")
+        save_checkpoint(unet,"fp16")
     if use_comet_ml:
         comet_experiment.end()
 accelerator.free_memory()

train-Copy1.py → train_chatgpt.py

@@ -7,12 +7,11 @@ 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
+from diffusers import UNet2DConditionModel, AutoencoderKL
 from accelerate import Accelerator
 from datasets import load_from_disk
 from tqdm import tqdm
-from PIL import Image,ImageOps
+from PIL import Image, ImageOps
 import wandb
 import random
 import gc
@@ -26,17 +25,17 @@ import torch.nn.functional as F
 from collections import deque
 
 # --------------------------- Параметры ---------------------------
-ds_path = "/workspace/sdxs3d/datasets/mjnj"
+ds_path = "/workspace/sdxs3d/datasets/butterfly"
 project = "unet"
 batch_size = 128
 base_learning_rate = 9e-5
 min_learning_rate = 1e-5
-num_epochs = 84
+num_epochs = 100
 # samples/save per epoch
-sample_interval_share = 5
+sample_interval_share = 1
 use_wandb = True
 use_comet_ml = False
-save_model = True
+save_model = False
 use_decay = True
 fbp = False # fused backward pass
 optimizer_type = "adam8bit"
@@ -45,8 +44,8 @@ unet_gradient = True
 clip_sample = False #Scheduler
 fixed_seed = True
 shuffle = True
-comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"  # Добавлен API ключ для Comet ML
-comet_ml_workspace = "recoilme"  # Добавлен workspace для Comet ML
+comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"
+comet_ml_workspace = "recoilme"
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
@@ -60,7 +59,7 @@ clip_grad_norm = 1.0
 steps_offset = 0 # Scheduler
 limit = 0
 checkpoints_folder = ""
-mixed_precision = "no" #"fp16"
+mixed_precision = "no" # "fp16"
 gradient_accumulation_steps = 1
 accelerator = Accelerator(
     mixed_precision=mixed_precision,
@@ -88,7 +87,7 @@ if fixed_seed:
         torch.cuda.manual_seed_all(seed)
 
 # --------------------------- Параметры LoRA ---------------------------
-lora_name = "" 
+lora_name = ""
 lora_rank = 32
 lora_alpha = 64
 
@@ -103,13 +102,6 @@ def sample_timesteps_bias(
     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":
@@ -126,17 +118,16 @@ def sample_timesteps_bias(
     timesteps = steps_offset + (samples * max_idx).long().to(device)
     return timesteps
 
+
 def logit_normal_samples(shape, mu=0.0, sigma=1.0, device=None, dtype=None):
     normal_samples = torch.normal(mean=mu, std=sigma, size=shape, device=device, dtype=dtype)
-
     logit_normal_samples = torch.sigmoid(normal_samples)
-
     return logit_normal_samples
 
 # --------------------------- Инициализация WandB ---------------------------
 if accelerator.is_main_process:
     if use_wandb:
-        wandb.init(project=project+lora_name, config={
+        wandb.init(project=project + lora_name, config={
             "batch_size": batch_size,
             "base_learning_rate": base_learning_rate,
             "num_epochs": num_epochs,
@@ -150,7 +141,6 @@ if accelerator.is_main_process:
             project_name=project,
             workspace=comet_ml_workspace
         )
-        # Логируем гиперпараметры в Comet ML
         hyper_params = {
             "batch_size": batch_size,
             "base_learning_rate": base_learning_rate,
@@ -170,8 +160,7 @@ 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 = AutoencoderKL.from_pretrained("AiArtLab/simplevae",subfolder="simple_vae_nightly",torch_dtype=dtype).to("cpu").eval()
+vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae", subfolder="vae", torch_dtype=dtype).to("cpu").eval()
 
 shift_factor = getattr(vae.config, "shift_factor", 0.0)
 if shift_factor is None:
@@ -185,8 +174,6 @@ latents_mean = getattr(vae.config, "latents_mean", None)
 latents_std = getattr(vae.config, "latents_std", None)
 
 
-
-
 class DistributedResolutionBatchSampler(Sampler):
     def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
         self.dataset = dataset
@@ -266,14 +253,16 @@ def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
     fixed_samples = {}
     for size, indices in size_groups.items():
         n_samples = min(samples_per_group, len(indices))
-        if len(size_groups)==1:
+        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)
+
+        # FIXED: keep fixed samples on CPU to avoid device/device-transfer issues when creating dataset
+        latents = torch.tensor(np.array([item["vae"] for item in samples_data]), dtype=dtype).cpu()
+        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data]), dtype=dtype).cpu()
         texts = [item["text"] for item in samples_data]
         fixed_samples[size] = (latents, embeddings, texts)
     
@@ -285,9 +274,10 @@ if limit > 0:
 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)
+    latents = torch.tensor(np.array([item["vae"] for item in batch]), dtype=dtype).to(device)
+    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch]), dtype=dtype).to(device)
     return latents, embeddings
 
 batch_sampler = DistributedResolutionBatchSampler(
@@ -298,20 +288,23 @@ batch_sampler = DistributedResolutionBatchSampler(
         shuffle=shuffle
     )
 
+# NOTE: we create dataloader first, then prepare it with accelerator. We'll create optimizer/lr_scheduler after
+# we know len(dataloader) (which is per-process after prepare) so that scheduling is consistent.
+
 dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
-print("Total samples",len(dataloader))
+print("Total batches (pre-prepare):", 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
+# --------------------------- Теперь безопасно --- вычисляем шаги и создаём оптимизатор/шедулер ---------------------------
+steps_per_epoch = len(dataloader)  # this is per-process (after prepare)
+total_training_steps = steps_per_epoch * num_epochs
+print(f"[rank {accelerator.process_index}] steps_per_epoch={steps_per_epoch}, total_training_steps={total_training_steps}")
 
-# Опция загрузки модели из последнего чекпоинта (если существует)
+# --------------------------- Создание/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)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device, dtype=dtype)
     if torch_compile:
         print("compiling")
         torch.set_float32_matmul_precision('high')
@@ -325,16 +318,14 @@ if os.path.isdir(latest_checkpoint):
     except Exception as e:
         print(f"Ошибка при включении SDPA: {e}")
         unet.set_use_memory_efficient_attention_xformers(True)
-
 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.")
 
+# --------------------------- LoRA (если нужно) ---------------------------
 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.")
 
@@ -372,8 +363,8 @@ if lora_name:
 if lora_name:
     trainable_params = [p for p in unet.parameters() if p.requires_grad]
 else:
-    if fbp:
-        trainable_params = list(unet.parameters())
+    trainable_params = list(unet.parameters()) if fbp else [p for p in unet.parameters() if p.requires_grad]
+
 
 def create_optimizer(name, params):
     if name == "adam8bit":
@@ -407,11 +398,15 @@ if fbp:
         optimizer_dict[param].zero_grad(set_to_none=True)
     for param in trainable_params:
         param.register_post_accumulate_grad_hook(optimizer_hook)
+    # FIXED: prepare fbp variant (keeps original logic)
     unet, optimizer = accelerator.prepare(unet, optimizer_dict)
 else:
-    optimizer = create_optimizer(optimizer_type, unet.parameters())
+    optimizer = create_optimizer(optimizer_type, trainable_params)
+
+    # FIXED: LR schedule should be based on total_training_steps (per-process steps * epochs)
     def lr_schedule(step):
-        x = step / (total_training_steps * world_size)
+        # step is current scheduler step (0..total_training_steps)
+        x = step / max(1, total_training_steps)
         warmup = warmup_percent
         if not use_decay:
             return base_learning_rate
@@ -420,6 +415,8 @@ else:
         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))
+
+    # LambdaLR expects a multiplier, so divide by base_learning_rate
     lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
 
     num_params = sum(p.numel() for p in unet.parameters())
@@ -427,12 +424,13 @@ else:
     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}")
+
+    # FIXED: prepare model, optimizer, scheduler AFTER creating them and after dataloader.prepare
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
 
 # --------------------------- 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
@@ -440,12 +438,12 @@ def generate_and_save_samples(fixed_samples_cpu, step):
         original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
         vae.to(device=device).eval()  # временно подгружаем VAE на GPU для декодинга
 
-        
         all_generated_images = []
         all_captions = [] 
         
         for size, (sample_latents, sample_text_embeddings, sample_text) in fixed_samples_cpu.items():
             width, height = size
+            # move CPU tensors to device here (they were kept on CPU in get_fixed_samples_by_resolution)
             sample_latents = sample_latents.to(dtype=dtype, device=device)
             sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
 
@@ -484,19 +482,11 @@ def generate_and_save_samples(fixed_samples_cpu, step):
 
             # Параметры нормализации
             latent_for_vae = current_latents.detach() / scaling_factor + shift_factor
-            
             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)
+                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"))
@@ -520,7 +510,7 @@ def generate_and_save_samples(fixed_samples_cpu, step):
                 wandb.Image(img, caption=f"{all_captions[i]}")
                 for i, img in enumerate(all_generated_images)
             ]
-            wandb.log({"generated_images": wandb_images})
+            wandb.log({"generated_images": wandb_images}, step=step)
         if use_comet_ml and accelerator.is_main_process:
             for i, img in enumerate(all_generated_images):
                 comet_experiment.log_image(
@@ -535,11 +525,7 @@ def generate_and_save_samples(fixed_samples_cpu, 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()
 
@@ -547,20 +533,22 @@ def generate_and_save_samples(fixed_samples_cpu, step):
 if accelerator.is_main_process:
     if save_model:
         print("Генерация сэмплов до старта обучения...")
-        generate_and_save_samples(fixed_samples,0)
+        generate_and_save_samples(fixed_samples, 0)
 accelerator.wait_for_everyone()
 
 # Модифицируем функцию сохранения модели для поддержки LoRA
-def save_checkpoint(unet,variant=""):
+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)
+            model_to_save = accelerator.unwrap_model(unet)
+            dest = os.path.join(checkpoints_folder, f"{project}")
+            if variant != "":
+                model_to_save.to(dtype=torch.float16).save_pretrained(dest, variant=variant)
             else:
-                accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
-            unet = unet.to(dtype=dtype)
+                model_to_save.save_pretrained(dest)
+
 
 # --------------------------- Тренировочный цикл ---------------------------
 if accelerator.is_main_process:
@@ -569,20 +557,26 @@ if accelerator.is_main_process:
 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 = 2.
+global_step = 0
+
+for epoch in range(0, num_epochs):
+    # FIXED: set epoch on the dataloader's batch_sampler if available (accelerator may have wrapped it)
+    if hasattr(dataloader, "batch_sampler") and hasattr(dataloader.batch_sampler, "set_epoch"):
+        dataloader.batch_sampler.set_epoch(epoch)
+    elif hasattr(batch_sampler, "set_epoch"):
+        batch_sampler.set_epoch(epoch)
 
-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()
-    #print("epoch:",epoch)
+
+    batch_losses = []
+    batch_grads = []
+
     for step, (latents, embeddings) in enumerate(dataloader):
         with accelerator.accumulate(unet):
-            if save_model == False and step == 5 :
+            if save_model == False and step == 5:
                 used_gb = torch.cuda.max_memory_allocated() / 1024**3
                 print(f"Шаг {step}: {used_gb:.2f} GB")
                 
@@ -596,80 +590,85 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
 
             mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
 
-            # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
-            batch_losses.append(mse_loss.detach().item())
+            # Сохраняем для локальных логов
+            batch_losses.append(mse_loss.detach().cpu().item())
 
-            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
-                accelerator.wait_for_everyone()
-                
             # Backward
             accelerator.backward(mse_loss)
 
-            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
-                accelerator.wait_for_everyone()
-                
-            grad = 0.0
+            grad_norm_val = 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)
+                    # Clip gradients and step only when gradients are synchronized (i.e. actual optimizer step)
+                    grad_norm = accelerator.clip_grad_norm_(unet.parameters(), clip_grad_norm)
+                    grad_norm_val = float(grad_norm)
                     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_grads.append(grad)
+            # increment global_step only when we have synchronized gradients (i.e. on optimizer step)
+            # FIXED: ensure global_step reflects optimizer updates, not micro-batches
+            if accelerator.sync_gradients:
+                global_step += 1
+                progress_bar.update(1)
+
+                # Логируем метрики только на главном процессе
+                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_norm_val)
+
+                    log_data = {
+                        "loss": mse_loss.detach().item(),
+                        "lr": current_lr,
+                        "grad": grad_norm_val,
+                        "epoch": epoch + 1,
+                        "global_step": global_step,
+                    }
 
-                log_data = {}
-                log_data["loss"] = 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)
 
-                # Генерируем сэмплы с заданным интервалом
-                if global_step % sample_interval == 0:
-                    generate_and_save_samples(fixed_samples,global_step)
-                    last_n = sample_interval
-
-                    if save_model:
-                        avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if len(batch_losses)>0 else 0.0
-                        print("saving:",avg_sample_loss < min_loss*save_barrier,"Current:",log_data["loss"],"Avg:",avg_sample_loss)
-                        if log_data["loss"] < min_loss*save_barrier:
-                            min_loss = avg_sample_loss
-                            save_checkpoint(unet)
-                    if use_wandb:
-                        wandb.log(log_data, step=global_step)
-                    if use_comet_ml:
-                        comet_experiment.log_metrics(log_data, step=global_step)
+                    # Генерируем сэмплы с заданным интервалом (только на главном процессе)
+                    if global_step % sample_interval == 0:
+                        generate_and_save_samples(fixed_samples, global_step)
 
+                        if save_model:
+                            avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if len(batch_losses) > 0 else 0.0
+                            if use_wandb:
+                                wandb.log({"sample_loss": avg_sample_loss})
+                            print("saving:", avg_sample_loss is not None and avg_sample_loss < min_loss * save_barrier, "Avg:", avg_sample_loss)
+                            if avg_sample_loss is not None and avg_sample_loss < min_loss * save_barrier:
+                                min_loss = avg_sample_loss
+                                save_checkpoint(unet)
 
+    # Эпоха окончена — агрегируем и логируем средние значения
     if accelerator.is_main_process:
-        avg_epoch_loss = np.mean(batch_losses[-steps_per_epoch:]) if len(batch_losses)>0 else 0.0
-        avg_epoch_grad = np.mean(batch_grads[-steps_per_epoch:]) if len(batch_grads)>0 else 0.0
+        # local averages
+        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({"epoch_loss": avg_epoch_loss, "epoch_grad": avg_epoch_grad, "epoch": epoch+1})
-        #if use_comet_ml:
-        #    comet_experiment.log_metrics(epoch_data)
+            wandb.log(log_data_ep)
+        if use_comet_ml:
+            comet_experiment.log_metrics(log_data_ep)
 
 # Завершение обучения - сохраняем финальную модель
 if accelerator.is_main_process:
     print("Обучение завершено! Сохраняем финальную модель...")
     if save_model:
-        save_checkpoint(unet,"fp16")
+        save_checkpoint(unet, "fp16")
     if use_comet_ml:
         comet_experiment.end()
 accelerator.free_memory()

{unet37ep → unet}/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e6fbf2b837fc99c48fa0815289cf1babe32a45dc7b975c4d0fcdcf41a7547367
+oid sha256:ad73d05c1584296954b3b137600b58f87c376f31cb5a80211a0dad2e3e118e9e
 size 6184944280

vae/config.json

@@ -0,0 +1,38 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.35.1",
+  "_name_or_path": "AiArtLab/simplevae",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": false,
+  "in_channels": 3,
+  "latent_channels": 16,
+  "latents_mean": null,
+  "latents_std": null,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 1024,
+  "scaling_factor": 1.0,
+  "shift_factor": 0,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": true,
+  "use_quant_conv": true
+}

vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ff1aa1cce9dcf003c7eb23548fdb62101fe9bdeedf30aa4a897b91a6ef13f93
+size 167669678