flow

train.py

@@ -8,7 +8,7 @@ from torch.utils.data.distributed import DistributedSampler
 from torch.optim.lr_scheduler import LambdaLR
 from collections import defaultdict
 from torch.optim.lr_scheduler import LambdaLR
-from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL, DDPMScheduler
+from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL
 from accelerate import Accelerator
 from datasets import load_from_disk
 from tqdm import tqdm
@@ -28,10 +28,10 @@ from collections import deque
 # --------------------------- Параметры ---------------------------
 ds_path = "/workspace/sdxs3d/datasets/butterfly"
 project = "unet"
-batch_size = 8
+batch_size = 16
 base_learning_rate = 9e-5
 min_learning_rate = 1e-5
-num_epochs = 300
+num_epochs = 30
 # samples/save per epoch
 sample_interval_share = 1
 use_wandb = True
@@ -42,16 +42,17 @@ optimizer_type = "adam8bit"
 torch_compile = False
 unet_gradient = True
 clip_sample = False #Scheduler
-fixed_seed = False
+fixed_seed = True
 shuffle = True
+use_comet_ml = False  # Добавлен флаг для Comet ML
+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)
 dtype = torch.float32
 save_barrier = 1.03
 warmup_percent = 0.01
-dispersive_temperature=0.5
-dispersive_weight= 0.05
 percentile_clipping = 99 # 8bit optim
 betta2 = 0.995
 eps = 1e-8
@@ -86,16 +87,6 @@ if fixed_seed:
     if torch.cuda.is_available():
         torch.cuda.manual_seed_all(seed)
 
-# --- Пропорции лоссов и окно медианного нормирования (КОЭФ., не значения) ---
-# CHANGED: добавлен huber и dispersive в пропорции, суммы = 1.0
-loss_ratios = {
-    "mse":   1.0,
-    "mae":   0.0,
-    "huber": 0.0,
-    "dispersive": 0.0,
-}
-median_coeff_steps = 128  # за сколько шагов считать медианные коэффициенты
-
 # --------------------------- Параметры LoRA ---------------------------
 lora_name = "" 
 lora_rank = 32
@@ -110,7 +101,7 @@ def sample_timesteps_bias(
     num_train_timesteps: int,  # обычно 1000
     steps_offset: int = 0,
     device=None,
-    mode: str = "uniform",        # "beta", "uniform"
+    mode: str = "beta",        # "beta", "uniform"
 ) -> torch.Tensor:
     """
     Возвращает timesteps с разным bias:
@@ -135,98 +126,45 @@ 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
+
 
-# Нормализация лоссов по медианам: считаем КОЭФФИЦИЕНТЫ
-class MedianLossNormalizer:
-    def __init__(self, desired_ratios: dict, window_steps: int):
-        # нормируем доли на случай, если сумма != 1
-        s = sum(desired_ratios.values())
-        self.ratios = {k: (v / s) for k, v in desired_ratios.items()}
-        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
-        self.window = window_steps
-
-    def update_and_total(self, losses: dict):
-        """
-        losses: dict ключ->тензор (значения лоссов)
-        Поведение:
-          - буферим ABS(l) только для активных (ratio>0) лоссов
-          - coeff = ratio / median(abs(loss))
-          - total = sum(coeff * loss) по активным лоссам
-        CHANGED: буферим abs() — чтобы медиана была положительной и не ломала деление.
-        """
-        # буферим только активные лоссы
-        for k, v in losses.items():
-            if k in self.buffers and self.ratios.get(k, 0) > 0:
-                self.buffers[k].append(float(v.detach().abs().cpu()))
-
-        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
-        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
-
-        # суммируем только по активным (ratio>0)
-        total = sum(coeffs[k] * losses[k] for k in coeffs if self.ratios.get(k, 0) > 0)
-        return total, coeffs, meds
-
-# создаём normalizer после определения loss_ratios
-normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
-
-class AccelerateDispersiveLoss:
-    def __init__(self, accelerator, temperature=0.5, weight=0.5):
-        self.accelerator = accelerator
-        self.temperature = temperature
-        self.weight = weight
-        self.activations = []
-        self.hooks = []
-    
-    def register_hooks(self, model, target_layer="down_blocks.0"):
-        unwrapped_model = self.accelerator.unwrap_model(model)
-        print("=== Поиск слоев в unwrapped модели ===")
-        for name, module in unwrapped_model.named_modules():
-            if target_layer in name:
-                hook = module.register_forward_hook(self.hook_fn)
-                self.hooks.append(hook)
-                print(f"✅ Хук зарегистрирован на: {name}")
-                break
-    
-    def hook_fn(self, module, input, output):
-        if isinstance(output, tuple):
-            activation = output[0]
-        else:
-            activation = output
-        if len(activation.shape) > 2:
-            activation = activation.view(activation.shape[0], -1)
-        self.activations.append(activation.detach().clone())
-        
-    def compute_dispersive_loss(self):
-        if not self.activations:
-            return torch.tensor(0.0, requires_grad=True, device=device)
-        local_activations = self.activations[-1].float()
-        batch_size = local_activations.shape[0]
-        if batch_size < 2:
-            return torch.tensor(0.0, requires_grad=True, device=device)
-        sf = local_activations / torch.norm(local_activations, dim=1, keepdim=True)
-        distance = torch.nn.functional.pdist(sf.float(), p=2) ** 2
-        exp_neg_dist = torch.exp(-distance / self.temperature) + 1e-5
-        dispersive_loss = torch.log(torch.mean(exp_neg_dist))
-        return dispersive_loss
-    
-    def clear_activations(self):
-        self.activations.clear()
-    
-    def remove_hooks(self):
-        for hook in self.hooks:
-            hook.remove()
-        self.hooks.clear()
 
 
 # --------------------------- Инициализация WandB ---------------------------
-if use_wandb and accelerator.is_main_process:
-    wandb.init(project=project+lora_name, config={
-        "batch_size": batch_size,
-        "base_learning_rate": base_learning_rate,
-        "num_epochs": num_epochs,
-        "fbp": fbp,
-        "optimizer_type": optimizer_type,
-    })
+if accelerator.is_main_process:
+    if use_wandb:
+        wandb.init(project=project+lora_name, config={
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "num_epochs": num_epochs,
+            "fbp": fbp,
+            "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
+        )
+        # Логируем гиперпараметры в Comet ML
+        hyper_params = {
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "min_learning_rate": min_learning_rate,
+            "num_epochs": num_epochs,
+            "n_diffusion_steps": n_diffusion_steps,
+            "guidance_scale": guidance_scale,
+            "optimizer_type": optimizer_type,
+            "mixed_precision": mixed_precision,
+        }
+        comet_experiment.log_parameters(hyper_params)
 
 # Включение Flash Attention 2/SDPA
 torch.backends.cuda.enable_flash_sdp(True)
@@ -236,12 +174,7 @@ gen.manual_seed(seed)
 
 # --------------------------- Загрузка моделей ---------------------------
 # VAE загружается на CPU для экономии GPU-памяти (как в твоём оригинальном коде)
-#vae = AutoencoderKLWan.from_pretrained("vae", variant="fp16").to(device="cpu", dtype=torch.float16).eval()
-#vae = AutoencoderKLWan.from_pretrained(
-#        "AiArtLab/simplevae", subfolder="wan16x_vae_nightly",
-#        torch_dtype=dtype
-#    ).to(device="cpu").eval()
-vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae",subfolder="simple_vae_nightly",torch_dtype=dtype).to(device).eval()
+vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae",subfolder="simple_vae_nightly",torch_dtype=dtype).to("cpu").eval()
 
 shift_factor = getattr(vae.config, "shift_factor", 0.0)
 if shift_factor is None:
@@ -254,14 +187,7 @@ if scaling_factor is None:
 latents_mean = getattr(vae.config, "latents_mean", None)
 latents_std = getattr(vae.config, "latents_std", None)
 
-# DDPMScheduler с V_Prediction и Zero-SNR
-scheduler = DDPMScheduler(
-    num_train_timesteps=1000,
-    prediction_type="v_prediction",
-    rescale_betas_zero_snr=True,
-    clip_sample = clip_sample,
-    steps_offset = steps_offset
-)
+
 
 
 class DistributedResolutionBatchSampler(Sampler):
@@ -403,13 +329,6 @@ if os.path.isdir(latest_checkpoint):
         print(f"Ошибка при включении SDPA: {e}")
         unet.set_use_memory_efficient_attention_xformers(True)
 
-    # Создаём hook для dispersive только если нужно
-    if loss_ratios.get("dispersive", 0) > 0:
-        dispersive_hook = AccelerateDispersiveLoss(
-            accelerator=accelerator,
-            temperature=dispersive_temperature,
-            weight=dispersive_weight
-        )
 else:
     # FIX: если чекпоинта нет — прекращаем с понятной ошибкой (лучше, чем неожиданные NameError дальше)
     raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}. Положи UNet чекпоинт в {latest_checkpoint} или укажи другой путь.")
@@ -512,10 +431,6 @@ else:
         if torch.isnan(param).any() or torch.isinf(param).any():
             print(f"[rank {accelerator.process_index}] NaN/Inf in {name}")
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
-    
-# Регистрация хуков ПОСЛЕ prepare
-if loss_ratios.get("dispersive", 0) > 0:
-    dispersive_hook.register_hooks(unet, "down_blocks.2")
 
 # --------------------------- Фиксированные семплы для генерации ---------------------------
 fixed_samples = get_fixed_samples_by_resolution(dataset)
@@ -528,7 +443,6 @@ 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
 
-        scheduler.set_timesteps(n_diffusion_steps)
         
         all_generated_images = []
         all_captions = [] 
@@ -546,39 +460,33 @@ def generate_and_save_samples(fixed_samples_cpu, step):
             )
             current_latents = noise.clone()
             
-            if guidance_scale > 0:
+            if guidance_scale != 1:
                 empty_embeddings = torch.zeros_like(sample_text_embeddings, dtype=sample_text_embeddings.dtype, device=device)
                 text_embeddings_batch = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
             else:
                 text_embeddings_batch = sample_text_embeddings
             
-            for t in scheduler.timesteps:
-                t_batch = t.repeat(current_latents.shape[0]).to(device)
-                if guidance_scale > 0:
-                    latent_model_input = torch.cat([current_latents] * 2)
+            timesteps = torch.linspace(0, 1, n_diffusion_steps+1, device=device, dtype=sample_latents.dtype)
+            for i in range(0, n_diffusion_steps):
+                t_cur = timesteps[i].unsqueeze(0)
+                t_next = timesteps[i+1]
+                dt = t_next - t_cur
+                if guidance_scale != 1:
+                    latent_model_input = torch.cat((current_latents, current_latents))
                 else:
                     latent_model_input = current_latents
+                t_batch = t_cur.repeat(latent_model_input.shape[0]).to(device)
+                t_batch = (t_batch * 1000).long().view(-1)
+                flow = original_model(latent_model_input, t_batch, text_embeddings_batch).sample               
                 
-                latent_model_input_scaled = scheduler.scale_model_input(latent_model_input, t_batch)
-                noise_pred = original_model(latent_model_input_scaled, t_batch, text_embeddings_batch).sample               
-                
-                if guidance_scale > 0:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+                if guidance_scale != 1:
+                    flow_uncond, flow_cond = flow.chunk(2)
+                    flow = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
 
-                current_latents = scheduler.step(noise_pred, t, current_latents).prev_sample
-
-            #print(current_latents.ndim, current_latents.shape)
-            #if current_latents.ndim == 4:
-            #    current_latents = current_latents.unsqueeze(2)
-            # Латент  в форме [B, C, T, H, W]
-            #print(current_latents.ndim, current_latents.shape)
+                current_latents = current_latents + flow * dt.to(device)
 
             # Параметры нормализации
-            latent_for_vae = current_latents.detach() * scaling_factor + shift_factor
-
-            if latents_mean!=None and latents_std!=None:
-                latent_for_vae = latent_for_vae * torch.tensor(latents_std, device=device, dtype=dtype).view(1, -1, 1, 1, 1) + torch.tensor(latents_mean, device=device, dtype=dtype).view(1, -1, 1, 1, 1)
+            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]
@@ -615,7 +523,20 @@ 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, "global_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(
+                    image_data=img,
+                    name=f"step_{step}_img_{i}",
+                    step=step,
+                    metadata={
+                        "caption": all_captions[i],
+                        "width": img.width,
+                        "height": img.height,
+                        "global_step": step
+                    }
+                )
     finally:
         # вернуть VAE на CPU (как было в твоём коде)
         vae.to("cpu")
@@ -644,30 +565,6 @@ def save_checkpoint(unet,variant=""):
                 accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
             unet = unet.to(dtype=dtype)
 
-def batch_pred_original_from_step(model_outputs, timesteps_tensor, noisy_latents, scheduler):
-    device = noisy_latents.device
-    dtype = noisy_latents.dtype
-
-    available_ts = scheduler.timesteps
-    if not isinstance(available_ts, torch.Tensor):
-        available_ts = torch.tensor(available_ts, device="cpu")
-    else:
-        available_ts = available_ts.cpu()
-
-    B = model_outputs.shape[0]
-    preds = []
-    for i in range(B):
-        t_i = int(timesteps_tensor[i].item())
-        diffs = torch.abs(available_ts - t_i)
-        idx = int(torch.argmin(diffs).item())
-        t_for_step = int(available_ts[idx].item())
-        model_out_i = model_outputs[i:i+1]
-        noisy_latent_i = noisy_latents[i:i+1]
-        step_out = scheduler.step(model_out_i, t_for_step, noisy_latent_i)
-        preds.append(step_out.pred_original_sample)
-
-    return torch.cat(preds, dim=0).to(device=device, dtype=dtype)
-
 # --------------------------- Тренировочный цикл ---------------------------
 if accelerator.is_main_process:
     print(f"Total steps per GPU: {total_training_steps}")
@@ -681,12 +578,11 @@ min_loss = 1.
 
 for epoch in range(start_epoch, start_epoch + num_epochs):
     batch_losses = []
-    batch_tlosses = []
     batch_grads = []
     batch_sampler.set_epoch(epoch)
     accelerator.wait_for_everyone()
     unet.train()
-    print("epoch:",epoch)
+    #print("epoch:",epoch)
     for step, (latents, embeddings) in enumerate(dataloader):
         with accelerator.accumulate(unet):
             if save_model == False and step == 5 :
@@ -695,44 +591,15 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
 
             noise = torch.randn_like(latents, dtype=latents.dtype)
 
-            progress = global_step / max(1, total_training_steps - 1)
-            timesteps = sample_timesteps_bias(
-                batch_size=latents.shape[0],
-                progress=progress,
-                num_train_timesteps=scheduler.config.num_train_timesteps,
-                steps_offset=steps_offset,
-                device=device
-            )
-
-            noisy_latents = scheduler.add_noise(latents, noise, timesteps)
-
-            if loss_ratios.get("dispersive", 0) > 0:
-                dispersive_hook.clear_activations()
+            t = logit_normal_samples((batch_size, 1, 1, 1), mu=0.0, sigma=1.0, device=latents.device, dtype=latents.dtype)
 
-            #print(latents.shape,embeddings.shape)
-            model_pred = unet(noisy_latents, timesteps, embeddings).sample
-            target_pred = scheduler.get_velocity(latents, noise, timesteps)
+            noisy_latents = (1 - t) * noise + t * latents
 
-            # === Losses ===
-            losses_dict = {}
+            t_for_unet = (t * 1000).long().view(-1)
+            model_pred = unet(noisy_latents, t_for_unet, embeddings).sample
+            target_pred = latents - noise
 
             mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
-            losses_dict["mse"] = mse_loss
-            losses_dict["mae"] = F.l1_loss(model_pred.float(), target_pred.float())
-
-            # CHANGED: Huber (smooth_l1) loss added
-            losses_dict["huber"] = F.smooth_l1_loss(model_pred.float(), target_pred.float())
-
-            # === Dispersive loss ===
-            if loss_ratios.get("dispersive", 0) > 0:
-                disp_raw = dispersive_hook.compute_dispersive_loss().to(device)  # может быть отрицательным
-                losses_dict["dispersive"] = dispersive_hook.weight * disp_raw
-            else:
-                losses_dict["dispersive"] = torch.tensor(0.0, device=device)
-
-            # === Нормализация всех лоссов ===
-            abs_for_norm = {k: losses_dict.get(k, torch.tensor(0.0, device=device)) for k in normalizer.ratios.keys()}
-            total_loss, coeffs, meds = normalizer.update_and_total(abs_for_norm)
 
             # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
             batch_losses.append(mse_loss.detach().item())
@@ -741,7 +608,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                 accelerator.wait_for_everyone()
                 
             # Backward
-            accelerator.backward(total_loss)
+            accelerator.backward(mse_loss)
 
             if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                 accelerator.wait_for_everyone()
@@ -765,32 +632,29 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                     current_lr = base_learning_rate
                 else:
                     current_lr = lr_scheduler.get_last_lr()[0]
-                batch_tlosses.append(total_loss.detach().item())
                 batch_grads.append(grad)
 
-                # Логируем только активные лоссы (ratio>0)
-                active_keys = [k for k, v in loss_ratios.items() if v > 0]
                 log_data = {}
-                for k in active_keys:
-                    v = losses_dict.get(k, None)
-                    if v is None:
-                        continue
-                    log_data[f"loss/{k}"] = (v.item() if isinstance(v, torch.Tensor) else float(v))
-
-                log_data["loss/total"] = float(total_loss.item())
-                log_data["loss/lr"] = current_lr
-                for k, c in coeffs.items():
-                    log_data[f"coeff/{k}"] = float(c)
-                if use_wandb and accelerator.sync_gradients:
-                    wandb.log(log_data, step=global_step)
+                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
                     avg_loss = float(np.mean(batch_losses[-last_n:])) if len(batch_losses) > 0 else 0.0
-                    avg_tloss = float(np.mean(batch_tlosses[-last_n:])) if len(batch_tlosses) > 0 else 0.0
                     avg_grad = float(np.mean(batch_grads[-last_n:])) if len(batch_grads) > 0 else 0.0
+
+                    avg_data = {}
+                    avg_data["avg_loss"] = avg_loss
+                    avg_data["avg_grad"] = avg_grad
+                    
                     print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, grad: {avg_grad:.6f}")
 
                     if save_model:
@@ -799,25 +663,26 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                             min_loss = avg_loss
                             save_checkpoint(unet)
                     if use_wandb:
-                        avg_data = {}
-                        avg_data["avg/loss"] = avg_loss
-                        avg_data["avg/tloss"] = avg_tloss
-                        avg_data["avg/grad"] = avg_grad
                         wandb.log(avg_data, step=global_step)
+                    if use_comet_ml:
+                        comet_experiment.log_metrics(avg_data, step=global_step)
+
 
     if accelerator.is_main_process:
         avg_epoch_loss = np.mean(batch_losses) if len(batch_losses)>0 else 0.0
         print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
         if use_wandb:
             wandb.log({"epoch_loss": avg_epoch_loss, "epoch": epoch+1})
+        #if use_comet_ml:
+        #    comet_experiment.log_metrics(epoch_data)
 
 # Завершение обучения - сохраняем финальную модель
-if loss_ratios.get("dispersive", 0) > 0:
-    dispersive_hook.remove_hooks()
 if accelerator.is_main_process:
     print("Обучение завершено! Сохраняем финальную 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()

train_velocity.py

@@ -0,0 +1,825 @@
+import os
+import math
+import torch
+import numpy as np
+import matplotlib.pyplot as plt
+from torch.utils.data import DataLoader, Sampler
+from torch.utils.data.distributed import DistributedSampler
+from torch.optim.lr_scheduler import LambdaLR
+from collections import defaultdict
+from torch.optim.lr_scheduler import LambdaLR
+from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL, DDPMScheduler
+from accelerate import Accelerator
+from datasets import load_from_disk
+from tqdm import tqdm
+from PIL import Image,ImageOps
+import wandb
+import random
+import gc
+from accelerate.state import DistributedType
+from torch.distributed import broadcast_object_list
+from torch.utils.checkpoint import checkpoint
+from diffusers.models.attention_processor import AttnProcessor2_0
+from datetime import datetime
+import bitsandbytes as bnb
+import torch.nn.functional as F
+from collections import deque
+
+# --------------------------- Параметры ---------------------------
+ds_path = "/workspace/sdxs3d/datasets/butterfly"
+project = "unet"
+batch_size = 16
+base_learning_rate = 9e-5
+min_learning_rate = 1e-5
+num_epochs = 300
+# samples/save per epoch
+sample_interval_share = 1
+use_wandb = True
+save_model = True
+use_decay = True
+fbp = False # fused backward pass
+optimizer_type = "adam8bit"
+torch_compile = False
+unet_gradient = True
+clip_sample = False #Scheduler
+fixed_seed = True
+shuffle = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+torch.backends.cuda.enable_mem_efficient_sdp(False)
+dtype = torch.float32
+save_barrier = 2.03 # TODO
+warmup_percent = 0.01
+dispersive_temperature=0.5
+dispersive_weight= 0.05
+percentile_clipping = 99 # 8bit optim
+betta2 = 0.995
+eps = 1e-8
+clip_grad_norm = 1.0
+steps_offset = 0 # Scheduler
+limit = 0
+checkpoints_folder = ""
+mixed_precision = "no" #"fp16"
+gradient_accumulation_steps = 1
+accelerator = Accelerator(
+    mixed_precision=mixed_precision,
+    gradient_accumulation_steps=gradient_accumulation_steps
+)
+device = accelerator.device
+
+# Параметры для диффузии
+n_diffusion_steps = 50
+samples_to_generate = 12
+guidance_scale = 5
+
+# Папки для сохранения результатов
+generated_folder = "samples"
+os.makedirs(generated_folder, exist_ok=True)
+
+# Настройка seed для воспроизводимости
+current_date = datetime.now()
+seed = int(current_date.strftime("%Y%m%d"))
+if fixed_seed:
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+# --- Пропорции лоссов и окно медианного нормирования (КОЭФ., не значения) ---
+# CHANGED: добавлен huber и dispersive в пропорции, суммы = 1.0
+loss_ratios = {
+    "mse":   1.0,
+    "mae":   0.0,
+    "huber": 0.0,
+    "dispersive": 0.0,
+}
+median_coeff_steps = 128  # за сколько шагов считать медианные коэффициенты
+
+# --------------------------- Параметры LoRA ---------------------------
+lora_name = "" 
+lora_rank = 32
+lora_alpha = 64
+
+print("init")
+
+# ---------------------------  вспомогательные функции ---------------------------
+def sample_timesteps_bias(
+    batch_size: int,
+    progress: float,           # [0..1]
+    num_train_timesteps: int,  # обычно 1000
+    steps_offset: int = 0,
+    device=None,
+    mode: str = "uniform",        # "beta", "uniform"
+) -> torch.Tensor:
+    """
+    Возвращает timesteps с разным bias:
+    - beta   : как раньше (сдвиг в начало или конец в зависимости от progress)
+    - normal : около середины (гауссовое распределение)
+    - uniform: равномерно по всем timestep’ам
+    """
+
+    max_idx = num_train_timesteps - 1 - steps_offset
+
+    if mode == "beta":
+        alpha = 1.0 + .5 * (1.0 - progress)
+        beta  = 1.0 + .5 * progress
+        samples = torch.distributions.Beta(alpha, beta).sample((batch_size,))
+
+    elif mode == "uniform":
+        samples = torch.rand(batch_size)
+
+    else:
+        raise ValueError(f"Unknown mode: {mode}")
+
+    timesteps = steps_offset + (samples * max_idx).long().to(device)
+    return timesteps
+
+
+# Нормализация лоссов по медианам: считаем КОЭФФИЦИЕНТЫ
+class MedianLossNormalizer:
+    def __init__(self, desired_ratios: dict, window_steps: int):
+        # нормируем доли на случай, если сумма != 1
+        s = sum(desired_ratios.values())
+        self.ratios = {k: (v / s) for k, v in desired_ratios.items()}
+        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        self.window = window_steps
+
+    def update_and_total(self, losses: dict):
+        """
+        losses: dict ключ->тензор (значения лоссов)
+        Поведение:
+          - буферим ABS(l) только для активных (ratio>0) лоссов
+          - coeff = ratio / median(abs(loss))
+          - total = sum(coeff * loss) по активным лоссам
+        CHANGED: буферим abs() — чтобы медиана была положительной и не ломала деление.
+        """
+        # буферим только активные лоссы
+        for k, v in losses.items():
+            if k in self.buffers and self.ratios.get(k, 0) > 0:
+                self.buffers[k].append(float(v.detach().abs().cpu()))
+
+        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
+        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
+
+        # суммируем только по активным (ratio>0)
+        total = sum(coeffs[k] * losses[k] for k in coeffs if self.ratios.get(k, 0) > 0)
+        return total, coeffs, meds
+
+# создаём normalizer после определения loss_ratios
+normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
+
+class AccelerateDispersiveLoss:
+    def __init__(self, accelerator, temperature=0.5, weight=0.5):
+        self.accelerator = accelerator
+        self.temperature = temperature
+        self.weight = weight
+        self.activations = []
+        self.hooks = []
+    
+    def register_hooks(self, model, target_layer="down_blocks.0"):
+        unwrapped_model = self.accelerator.unwrap_model(model)
+        print("=== Поиск слоев в unwrapped модели ===")
+        for name, module in unwrapped_model.named_modules():
+            if target_layer in name:
+                hook = module.register_forward_hook(self.hook_fn)
+                self.hooks.append(hook)
+                print(f"✅ Хук зарегистрирован на: {name}")
+                break
+    
+    def hook_fn(self, module, input, output):
+        if isinstance(output, tuple):
+            activation = output[0]
+        else:
+            activation = output
+        if len(activation.shape) > 2:
+            activation = activation.view(activation.shape[0], -1)
+        self.activations.append(activation.detach().clone())
+        
+    def compute_dispersive_loss(self):
+        if not self.activations:
+            return torch.tensor(0.0, requires_grad=True, device=device)
+        local_activations = self.activations[-1].float()
+        batch_size = local_activations.shape[0]
+        if batch_size < 2:
+            return torch.tensor(0.0, requires_grad=True, device=device)
+        sf = local_activations / torch.norm(local_activations, dim=1, keepdim=True)
+        distance = torch.nn.functional.pdist(sf.float(), p=2) ** 2
+        exp_neg_dist = torch.exp(-distance / self.temperature) + 1e-5
+        dispersive_loss = torch.log(torch.mean(exp_neg_dist))
+        return dispersive_loss
+    
+    def clear_activations(self):
+        self.activations.clear()
+    
+    def remove_hooks(self):
+        for hook in self.hooks:
+            hook.remove()
+        self.hooks.clear()
+
+
+# --------------------------- Инициализация WandB ---------------------------
+if use_wandb and accelerator.is_main_process:
+    wandb.init(project=project+lora_name, config={
+        "batch_size": batch_size,
+        "base_learning_rate": base_learning_rate,
+        "num_epochs": num_epochs,
+        "fbp": fbp,
+        "optimizer_type": optimizer_type,
+    })
+
+# Включение Flash Attention 2/SDPA
+torch.backends.cuda.enable_flash_sdp(True)
+# --------------------------- Инициализация Accelerator --------------------
+gen = torch.Generator(device=device)
+gen.manual_seed(seed)
+
+# --------------------------- Загрузка моделей ---------------------------
+# VAE загружается на CPU для экономии GPU-памяти (как в твоём оригинальном коде)
+#vae = AutoencoderKLWan.from_pretrained("vae", variant="fp16").to(device="cpu", dtype=torch.float16).eval()
+#vae = AutoencoderKLWan.from_pretrained(
+#        "AiArtLab/simplevae", subfolder="wan16x_vae_nightly",
+#        torch_dtype=dtype
+#    ).to(device="cpu").eval()
+vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae",subfolder="simple_vae_nightly",torch_dtype=dtype).to(device).eval()
+
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+latents_mean = getattr(vae.config, "latents_mean", None)
+latents_std = getattr(vae.config, "latents_std", None)
+
+# DDPMScheduler с V_Prediction и Zero-SNR
+scheduler = DDPMScheduler(
+    num_train_timesteps=1000,
+    prediction_type="v_prediction",
+    rescale_betas_zero_snr=True,
+    clip_sample = clip_sample,
+    steps_offset = steps_offset
+)
+
+
+class DistributedResolutionBatchSampler(Sampler):
+    def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
+        self.dataset = dataset
+        self.batch_size = max(1, batch_size // num_replicas)
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.shuffle = shuffle
+        self.drop_last = drop_last
+        self.epoch = 0
+        
+        try:
+            widths = np.array(dataset["width"])
+            heights = np.array(dataset["height"])
+        except KeyError:
+            widths = np.zeros(len(dataset))
+            heights = np.zeros(len(dataset))
+        
+        self.size_keys = np.unique(np.stack([widths, heights], axis=1), axis=0)
+        self.size_groups = {}
+        for w, h in self.size_keys:
+            mask = (widths == w) & (heights == h)
+            self.size_groups[(w, h)] = np.where(mask)[0]
+            
+        self.group_num_batches = {}
+        total_batches = 0
+        for size, indices in self.size_groups.items():
+            num_full_batches = len(indices) // (self.batch_size * self.num_replicas)
+            self.group_num_batches[size] = num_full_batches
+            total_batches += num_full_batches
+            
+        self.num_batches = (total_batches // self.num_replicas) * self.num_replicas
+        
+    def __iter__(self):
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        all_batches = []
+        rng = np.random.RandomState(self.epoch)
+        
+        for size, indices in self.size_groups.items():
+            indices = indices.copy()
+            if self.shuffle:
+                rng.shuffle(indices)
+            num_full_batches = self.group_num_batches[size]
+            if num_full_batches == 0:
+                continue
+            valid_indices = indices[:num_full_batches * self.batch_size * self.num_replicas]
+            batches = valid_indices.reshape(-1, self.batch_size * self.num_replicas)
+            start_idx = self.rank * self.batch_size
+            end_idx = start_idx + self.batch_size
+            gpu_batches = batches[:, start_idx:end_idx]
+            all_batches.extend(gpu_batches)
+        
+        if self.shuffle:
+            rng.shuffle(all_batches)
+        accelerator.wait_for_everyone()
+        return iter(all_batches)
+
+    def __len__(self):
+        return self.num_batches 
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+# Функция для выборки фиксированных семплов по размерам
+def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
+    size_groups = defaultdict(list)
+    try:
+        widths = dataset["width"]
+        heights = dataset["height"]
+    except KeyError:
+        widths = [0] * len(dataset)
+        heights = [0] * len(dataset)
+    for i, (w, h) in enumerate(zip(widths, heights)):
+        size = (w, h)
+        size_groups[size].append(i)
+    
+    fixed_samples = {}
+    for size, indices in size_groups.items():
+        n_samples = min(samples_per_group, len(indices))
+        if len(size_groups)==1:
+            n_samples = samples_to_generate
+        if n_samples == 0:
+            continue
+        sample_indices = random.sample(indices, n_samples)
+        samples_data = [dataset[idx] for idx in sample_indices]
+        latents = torch.tensor(np.array([item["vae"] for item in samples_data])).to(device=device,dtype=dtype)
+        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data])).to(device,dtype=dtype)
+        texts = [item["text"] for item in samples_data]
+        fixed_samples[size] = (latents, embeddings, texts)
+    
+    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям")
+    return fixed_samples
+
+if limit > 0:
+    dataset = load_from_disk(ds_path).select(range(limit))
+else:
+    dataset = load_from_disk(ds_path)
+
+def collate_fn_simple(batch):
+    latents = torch.tensor(np.array([item["vae"] for item in batch])).to(device,dtype=dtype)
+    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch])).to(device,dtype=dtype)
+    return latents, embeddings
+
+batch_sampler = DistributedResolutionBatchSampler(
+        dataset=dataset,
+        batch_size=batch_size,
+        num_replicas=accelerator.num_processes,
+        rank=accelerator.process_index,
+        shuffle=shuffle
+    )
+
+dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_simple)
+print("Total samples",len(dataloader))
+dataloader = accelerator.prepare(dataloader)
+
+start_epoch = 0
+global_step = 0
+total_training_steps = (len(dataloader) * num_epochs)
+world_size = accelerator.state.num_processes
+
+# Опция загрузки модели из последнего чекпоинта (если существует)
+latest_checkpoint = os.path.join(checkpoints_folder, project)
+if os.path.isdir(latest_checkpoint):
+    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
+    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device,dtype=dtype)
+    if torch_compile:
+        print("compiling")
+        torch.set_float32_matmul_precision('high')
+        unet = torch.compile(unet)
+        print("compiling - ok")
+    if unet_gradient:
+        unet.enable_gradient_checkpointing()
+    unet.set_use_memory_efficient_attention_xformers(False)
+    try:
+        unet.set_attn_processor(AttnProcessor2_0())
+    except Exception as e:
+        print(f"Ошибка при включении SDPA: {e}")
+        unet.set_use_memory_efficient_attention_xformers(True)
+
+    # Создаём hook для dispersive только если нужно
+    if loss_ratios.get("dispersive", 0) > 0:
+        dispersive_hook = AccelerateDispersiveLoss(
+            accelerator=accelerator,
+            temperature=dispersive_temperature,
+            weight=dispersive_weight
+        )
+else:
+    # FIX: если чекпоинта нет — прекращаем с понятной ошибкой (лучше, чем неожиданные NameError дальше)
+    raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}. Положи UNet чекпоинт в {latest_checkpoint} или укажи другой путь.")
+
+if lora_name:
+    print(f"--- Настройка LoRA через PEFT (Rank={lora_rank}, Alpha={lora_alpha}) ---")
+    from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
+    from peft.tuners.lora import LoraModel
+    import os
+    unet.requires_grad_(False)
+    print("Параметры базового UNet заморожены.")
+
+    lora_config = LoraConfig(
+        r=lora_rank,
+        lora_alpha=lora_alpha,
+        target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+    )
+    unet.add_adapter(lora_config)
+
+    from peft import get_peft_model
+    peft_unet = get_peft_model(unet, lora_config)
+    params_to_optimize = list(p for p in peft_unet.parameters() if p.requires_grad)
+
+    if accelerator.is_main_process:
+        lora_params_count = sum(p.numel() for p in params_to_optimize)
+        total_params_count = sum(p.numel() for p in unet.parameters())
+        print(f"Количество обучаемых параметров (LoRA): {lora_params_count:,}")
+        print(f"Общее количество параметров UNet: {total_params_count:,}")
+
+    lora_save_path = os.path.join("lora", lora_name)
+    os.makedirs(lora_save_path, exist_ok=True)
+
+    def save_lora_checkpoint(model):
+        if accelerator.is_main_process:
+            print(f"Сохраняем LoRA адаптеры  в {lora_save_path}")
+            from peft.utils.save_and_load import get_peft_model_state_dict
+            lora_state_dict = get_peft_model_state_dict(model)
+            torch.save(lora_state_dict, os.path.join(lora_save_path, "adapter_model.bin"))
+            model.peft_config["default"].save_pretrained(lora_save_path)
+            from diffusers import StableDiffusionXLPipeline
+            StableDiffusionXLPipeline.save_lora_weights(lora_save_path, lora_state_dict)
+
+# --------------------------- Оптимизатор ---------------------------
+if lora_name:
+    trainable_params = [p for p in unet.parameters() if p.requires_grad]
+else:
+    if fbp:
+        trainable_params = list(unet.parameters())
+
+def create_optimizer(name, params):
+    if name == "adam8bit":
+        return bnb.optim.AdamW8bit(
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01,
+            percentile_clipping=percentile_clipping
+        )
+    elif name == "adam":
+        return torch.optim.AdamW(
+            params, lr=base_learning_rate, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01
+        )
+    elif name == "lion8bit":
+        return bnb.optim.Lion8bit(
+            params, lr=base_learning_rate, betas=(0.9, 0.97), weight_decay=0.01,
+            percentile_clipping=percentile_clipping
+        )
+    elif name == "adafactor":
+        from transformers import Adafactor
+        return Adafactor(
+            params, lr=base_learning_rate, scale_parameter=True, relative_step=False,
+            warmup_init=False, eps=(1e-30, 1e-3), clip_threshold=1.0,
+            beta1=0.9, weight_decay=0.01
+        )
+    else:
+        raise ValueError(f"Unknown optimizer: {name}")
+
+if fbp:
+    optimizer_dict = {p: create_optimizer(optimizer_type, [p]) for p in trainable_params}
+    def optimizer_hook(param):
+        optimizer_dict[param].step()
+        optimizer_dict[param].zero_grad(set_to_none=True)
+    for param in trainable_params:
+        param.register_post_accumulate_grad_hook(optimizer_hook)
+    unet, optimizer = accelerator.prepare(unet, optimizer_dict)
+else:
+    optimizer = create_optimizer(optimizer_type, unet.parameters())
+    def lr_schedule(step):
+        x = step / (total_training_steps * world_size)
+        warmup = warmup_percent
+        if not use_decay:
+            return base_learning_rate
+        if x < warmup:
+            return min_learning_rate + (base_learning_rate - min_learning_rate) * (x / warmup)
+        decay_ratio = (x - warmup) / (1 - warmup)
+        return min_learning_rate + 0.5 * (base_learning_rate - min_learning_rate) * \
+               (1 + math.cos(math.pi * decay_ratio))
+    lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)
+
+    num_params = sum(p.numel() for p in unet.parameters())
+    print(f"[rank {accelerator.process_index}] total params: {num_params}")
+    for name, param in unet.named_parameters():
+        if torch.isnan(param).any() or torch.isinf(param).any():
+            print(f"[rank {accelerator.process_index}] NaN/Inf in {name}")
+    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
+    
+# Регистрация хуков ПОСЛЕ prepare
+if loss_ratios.get("dispersive", 0) > 0:
+    dispersive_hook.register_hooks(unet, "down_blocks.2")
+
+# --------------------------- Фиксированные семплы для генерации ---------------------------
+fixed_samples = get_fixed_samples_by_resolution(dataset)
+
+@torch.compiler.disable()
+@torch.no_grad()
+def generate_and_save_samples(fixed_samples_cpu, step):
+    original_model = None
+    try:
+        original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
+        vae.to(device=device).eval()  # временно подгружаем VAE на GPU для декодинга
+
+        scheduler.set_timesteps(n_diffusion_steps)
+        
+        all_generated_images = []
+        all_captions = [] 
+        
+        for size, (sample_latents, sample_text_embeddings, sample_text) in fixed_samples_cpu.items():
+            width, height = size
+            sample_latents = sample_latents.to(dtype=dtype, device=device)
+            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
+
+            noise = torch.randn(
+                sample_latents.shape,
+                generator=gen,
+                device=device,
+                dtype=sample_latents.dtype
+            )
+            current_latents = noise.clone()
+            
+            if guidance_scale > 0:
+                empty_embeddings = torch.zeros_like(sample_text_embeddings, dtype=sample_text_embeddings.dtype, device=device)
+                text_embeddings_batch = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
+            else:
+                text_embeddings_batch = sample_text_embeddings
+            
+            for t in scheduler.timesteps:
+                t_batch = t.repeat(current_latents.shape[0]).to(device)
+                if guidance_scale > 0:
+                    latent_model_input = torch.cat([current_latents] * 2)
+                else:
+                    latent_model_input = current_latents
+                
+                latent_model_input_scaled = scheduler.scale_model_input(latent_model_input, t_batch)
+                noise_pred = original_model(latent_model_input_scaled, t_batch, text_embeddings_batch).sample               
+                
+                if guidance_scale > 0:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                current_latents = scheduler.step(noise_pred, t, current_latents).prev_sample
+
+            #print(current_latents.ndim, current_latents.shape)
+            #if current_latents.ndim == 4:
+            #    current_latents = current_latents.unsqueeze(2)
+            # Латент  в форме [B, C, T, H, W]
+            #print(current_latents.ndim, current_latents.shape)
+
+            # Параметры нормализации
+            latent_for_vae = current_latents.detach() * scaling_factor + shift_factor
+
+            if latents_mean!=None and latents_std!=None:
+                latent_for_vae = latent_for_vae * torch.tensor(latents_std, device=device, dtype=dtype).view(1, -1, 1, 1, 1) + torch.tensor(latents_mean, device=device, dtype=dtype).view(1, -1, 1, 1, 1)
+            
+            decoded = vae.decode(latent_for_vae.to(torch.float32)).sample
+            #decoded = decoded[:, :, 0, :, :]  # [3, H, W]
+            #print(decoded.ndim, decoded.shape)
+            
+            decoded_fp32 = decoded.to(torch.float32)
+            for img_idx, img_tensor in enumerate(decoded_fp32):
+
+                # Форма: [3, H, W] -> преобразуем в [H, W, 3]
+                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
+                img = img.transpose(1, 2, 0)  # Из [3, H, W] в [H, W, 3]
+    
+                #img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
+                if np.isnan(img).any():
+                    print("NaNs found, saving stopped! Step:", step)
+                pil_img = Image.fromarray((img * 255).astype("uint8"))
+                
+                max_w_overall = max(s[0] for s in fixed_samples_cpu.keys())
+                max_h_overall = max(s[1] for s in fixed_samples_cpu.keys())
+                max_w_overall = max(255, max_w_overall)
+                max_h_overall = max(255, max_h_overall)
+            
+                padded_img = ImageOps.pad(pil_img, (max_w_overall, max_h_overall), color='white')
+                all_generated_images.append(padded_img)
+
+                caption_text = sample_text[img_idx][:200] if img_idx < len(sample_text) else ""
+                all_captions.append(caption_text)
+                
+                sample_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
+                pil_img.save(sample_path, "JPEG", quality=96)
+        
+        if use_wandb and accelerator.is_main_process:
+            wandb_images = [
+                wandb.Image(img, caption=f"{all_captions[i]}")
+                for i, img in enumerate(all_generated_images)
+            ]
+            wandb.log({"generated_images": wandb_images, "global_step": step})
+    finally:
+        # вернуть VAE на CPU (как было в твоём коде)
+        vae.to("cpu")
+        for var in list(locals().keys()):
+            if isinstance(locals()[var], torch.Tensor):
+                del locals()[var]
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# --------------------------- Генерация сэмплов перед обучением ---------------------------
+if accelerator.is_main_process:
+    if save_model:
+        print("Генерация сэмплов до старта обучения...")
+        generate_and_save_samples(fixed_samples,0)
+accelerator.wait_for_everyone()
+
+# Модифицируем функцию сохранения модели для поддержки LoRA
+def save_checkpoint(unet,variant=""):
+    if accelerator.is_main_process:
+        if lora_name:
+            save_lora_checkpoint(unet)
+        else:
+            if variant!="":
+                accelerator.unwrap_model(unet.to(dtype=torch.float16)).save_pretrained(os.path.join(checkpoints_folder, f"{project}"),variant=variant)
+            else:
+                accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
+            unet = unet.to(dtype=dtype)
+
+def batch_pred_original_from_step(model_outputs, timesteps_tensor, noisy_latents, scheduler):
+    device = noisy_latents.device
+    dtype = noisy_latents.dtype
+
+    available_ts = scheduler.timesteps
+    if not isinstance(available_ts, torch.Tensor):
+        available_ts = torch.tensor(available_ts, device="cpu")
+    else:
+        available_ts = available_ts.cpu()
+
+    B = model_outputs.shape[0]
+    preds = []
+    for i in range(B):
+        t_i = int(timesteps_tensor[i].item())
+        diffs = torch.abs(available_ts - t_i)
+        idx = int(torch.argmin(diffs).item())
+        t_for_step = int(available_ts[idx].item())
+        model_out_i = model_outputs[i:i+1]
+        noisy_latent_i = noisy_latents[i:i+1]
+        step_out = scheduler.step(model_out_i, t_for_step, noisy_latent_i)
+        preds.append(step_out.pred_original_sample)
+
+    return torch.cat(preds, dim=0).to(device=device, dtype=dtype)
+
+# --------------------------- Тренировочный цикл ---------------------------
+if accelerator.is_main_process:
+    print(f"Total steps per GPU: {total_training_steps}")
+
+epoch_loss_points = []
+progress_bar = tqdm(total=total_training_steps, disable=not accelerator.is_local_main_process, desc="Training", unit="step")
+
+steps_per_epoch = len(dataloader)
+sample_interval = max(1, steps_per_epoch // sample_interval_share)
+min_loss = 1.
+
+for epoch in range(start_epoch, start_epoch + num_epochs):
+    batch_losses = []
+    batch_tlosses = []
+    batch_grads = []
+    batch_sampler.set_epoch(epoch)
+    accelerator.wait_for_everyone()
+    unet.train()
+    print("epoch:",epoch)
+    for step, (latents, embeddings) in enumerate(dataloader):
+        with accelerator.accumulate(unet):
+            if save_model == False and step == 5 :
+                used_gb = torch.cuda.max_memory_allocated() / 1024**3
+                print(f"Шаг {step}: {used_gb:.2f} GB")
+
+            noise = torch.randn_like(latents, dtype=latents.dtype)
+
+            progress = global_step / max(1, total_training_steps - 1)
+            timesteps = sample_timesteps_bias(
+                batch_size=latents.shape[0],
+                progress=progress,
+                num_train_timesteps=scheduler.config.num_train_timesteps,
+                steps_offset=steps_offset,
+                device=device
+            )
+
+            noisy_latents = scheduler.add_noise(latents, noise, timesteps)
+
+            if loss_ratios.get("dispersive", 0) > 0:
+                dispersive_hook.clear_activations()
+
+            #print(latents.shape,embeddings.shape)
+            model_pred = unet(noisy_latents, timesteps, embeddings).sample
+            target_pred = scheduler.get_velocity(latents, noise, timesteps)
+
+            # === Losses ===
+            losses_dict = {}
+
+            mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
+            losses_dict["mse"] = mse_loss
+            losses_dict["mae"] = F.l1_loss(model_pred.float(), target_pred.float())
+
+            # CHANGED: Huber (smooth_l1) loss added
+            losses_dict["huber"] = F.smooth_l1_loss(model_pred.float(), target_pred.float())
+
+            # === Dispersive loss ===
+            if loss_ratios.get("dispersive", 0) > 0:
+                disp_raw = dispersive_hook.compute_dispersive_loss().to(device)  # может быть отрицательным
+                losses_dict["dispersive"] = dispersive_hook.weight * disp_raw
+            else:
+                losses_dict["dispersive"] = torch.tensor(0.0, device=device)
+
+            # === Нормализация всех лоссов ===
+            abs_for_norm = {k: losses_dict.get(k, torch.tensor(0.0, device=device)) for k in normalizer.ratios.keys()}
+            total_loss, coeffs, meds = normalizer.update_and_total(abs_for_norm)
+
+            # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
+            batch_losses.append(mse_loss.detach().item())
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+                
+            # Backward
+            accelerator.backward(total_loss)
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+                
+            grad = 0.0
+            if not fbp:
+                if accelerator.sync_gradients:
+                    with torch.amp.autocast('cuda', enabled=False):
+                        grad_val = accelerator.clip_grad_norm_(unet.parameters(), clip_grad_norm)
+                    grad = float(grad_val)
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+
+            global_step += 1
+            progress_bar.update(1)
+
+            # Логируем метрики
+            if accelerator.is_main_process:
+                if fbp:
+                    current_lr = base_learning_rate
+                else:
+                    current_lr = lr_scheduler.get_last_lr()[0]
+                batch_tlosses.append(total_loss.detach().item())
+                batch_grads.append(grad)
+
+                # Логируем только активные лоссы (ratio>0)
+                active_keys = [k for k, v in loss_ratios.items() if v > 0]
+                log_data = {}
+                for k in active_keys:
+                    v = losses_dict.get(k, None)
+                    if v is None:
+                        continue
+                    log_data[f"loss/{k}"] = (v.item() if isinstance(v, torch.Tensor) else float(v))
+
+                log_data["loss/total"] = float(total_loss.item())
+                log_data["loss/lr"] = current_lr
+                for k, c in coeffs.items():
+                    log_data[f"coeff/{k}"] = float(c)
+                if use_wandb and accelerator.sync_gradients:
+                    wandb.log(log_data, step=global_step)
+
+                # Генерируем сэмплы с заданным интервалом
+                if global_step % sample_interval == 0:
+                    generate_and_save_samples(fixed_samples,global_step)
+                    last_n = sample_interval
+                    avg_loss = float(np.mean(batch_losses[-last_n:])) if len(batch_losses) > 0 else 0.0
+                    avg_tloss = float(np.mean(batch_tlosses[-last_n:])) if len(batch_tlosses) > 0 else 0.0
+                    avg_grad = float(np.mean(batch_grads[-last_n:])) if len(batch_grads) > 0 else 0.0
+                    print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, grad: {avg_grad:.6f}")
+
+                    if save_model:
+                        print("saving:",avg_loss < min_loss*save_barrier)
+                        if avg_loss < min_loss*save_barrier:
+                            min_loss = avg_loss
+                            save_checkpoint(unet)
+                    if use_wandb:
+                        avg_data = {}
+                        avg_data["avg/loss"] = avg_loss
+                        avg_data["avg/tloss"] = avg_tloss
+                        avg_data["avg/grad"] = avg_grad
+                        wandb.log(avg_data, step=global_step)
+
+    if accelerator.is_main_process:
+        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses)>0 else 0.0
+        print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
+        if use_wandb:
+            wandb.log({"epoch_loss": avg_epoch_loss, "epoch": epoch+1})
+
+# Завершение обучения - сохраняем финальную модель
+if loss_ratios.get("dispersive", 0) > 0:
+    dispersive_hook.remove_hooks()
+if accelerator.is_main_process:
+    print("Обучение завершено! Сохраняем финальную модель...")
+    if save_model:
+        save_checkpoint(unet,"fp16")
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+    
+print("Готово!")

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:23e75f6769d03f650a4673248b33bfd2af15f07d4e8bc75bbc8eca6420f33a8d
+oid sha256:15dbbdb970577298a421e4424ea6ff535526b6369a73decc5eac24b2401487d6
 size 3092571208