Various code to support experiments.

toolkit/config_modules.py

@@ -437,7 +437,7 @@ class TrainConfig:
         # adds an additional loss to the network to encourage it output a normalized standard deviation
         self.target_norm_std = kwargs.get('target_norm_std', None)
         self.target_norm_std_value = kwargs.get('target_norm_std_value', 1.0)
-        self.timestep_type = kwargs.get('timestep_type', 'sigmoid')  # sigmoid, linear, lognorm_blend, next_sample, weighted
+        self.timestep_type = kwargs.get('timestep_type', 'sigmoid')  # sigmoid, linear, lognorm_blend, next_sample, weighted, one_step
         self.next_sample_timesteps = kwargs.get('next_sample_timesteps', 8)
         self.linear_timesteps = kwargs.get('linear_timesteps', False)
         self.linear_timesteps2 = kwargs.get('linear_timesteps2', False)

toolkit/models/diffusion_feature_extraction.py

@@ -1,3 +1,4 @@
+import math
 import torch
 import os
 from torch import nn
@@ -351,12 +352,251 @@ class DiffusionFeatureExtractor3(nn.Module):
         
         return total_loss
 
+class DiffusionFeatureExtractor4(nn.Module):
+    def __init__(self, device=torch.device("cuda"), dtype=torch.bfloat16, vae=None):
+        super().__init__()
+        self.version = 4
+        if vae is None:
+            raise ValueError("vae must be provided for DFE4")
+        self.vae = vae
+        # image_encoder_path = "google/siglip-so400m-patch14-384"
+        image_encoder_path = "google/siglip2-so400m-patch16-naflex"
+        from transformers import Siglip2ImageProcessor, Siglip2VisionModel
+        try:
+            self.image_processor = Siglip2ImageProcessor.from_pretrained(
+                image_encoder_path)
+        except EnvironmentError:
+            self.image_processor = Siglip2ImageProcessor()
+        
+        self.image_processor.max_num_patches = 1024
+            
+        self.vision_encoder = Siglip2VisionModel.from_pretrained(
+            image_encoder_path,
+            ignore_mismatched_sizes=True
+        ).to(device, dtype=dtype)
+
+        self.losses = {}
+        self.log_every = 100
+        self.step = 0
+    
+    def _target_hw(self, h, w, patch, max_patches, eps: float = 1e-5):
+        def _snap(x, s):
+            x = math.ceil((x * s) / patch) * patch
+            return max(patch, int(x))
+
+        lo, hi = eps / 10, 1.0
+        while hi - lo >= eps:
+            mid = (lo + hi) / 2
+            th, tw = _snap(h, mid), _snap(w, mid)
+            if (th // patch) * (tw // patch) <= max_patches:
+                lo = mid
+            else:
+                hi = mid
+        return _snap(h, lo), _snap(w, lo)
+
+
+    def tensors_to_siglip_like_features(self, batch: torch.Tensor):
+        """
+        Args:
+            batch: (bs, 3, H, W) tensor already in the desired value range
+                   (e.g. [-1, 1] or [0, 1]); no extra rescale / normalize here.
+
+        Returns:
+            dict(
+                pixel_values        – (bs, L, P) where L = n_h*n_w, P = 3*patch*patch
+                pixel_attention_mask– (L,)  all-ones
+                spatial_shapes      – (n_h, n_w)
+            )
+        """
+        if batch.ndim != 4:
+            raise ValueError("Expected (bs, 3, H, W) tensor")
+
+        bs, c, H, W = batch.shape
+        proc        = self.image_processor
+        patch       = proc.patch_size
+        max_patches = proc.max_num_patches
+
+        # One shared resize for the whole batch
+        tgt_h, tgt_w = self._target_hw(H, W, patch, max_patches)
+        batch = torch.nn.functional.interpolate(
+            batch, size=(tgt_h, tgt_w), mode="bilinear", align_corners=False
+        )
+
+        n_h, n_w = tgt_h // patch, tgt_w // patch
+        # flat_dim = c * patch * patch
+        num_p    = n_h * n_w
+
+        # unfold → (bs, flat_dim, num_p)  → (bs, num_p, flat_dim)
+        patches = (
+            torch.nn.functional.unfold(batch, kernel_size=patch, stride=patch)
+            .transpose(1, 2)
+        )
+
+        attn_mask = torch.ones(num_p, dtype=torch.long, device=batch.device)
+        spatial   = torch.tensor((n_h, n_w), device=batch.device, dtype=torch.int32)
+        
+        # repeat attn_mask for each batch element
+        attn_mask = attn_mask.unsqueeze(0).repeat(bs, 1)
+        spatial = spatial.unsqueeze(0).repeat(bs, 1)
+
+        return {
+            "pixel_values": patches,            # (bs, num_patches, patch_dim)
+            "pixel_attention_mask": attn_mask,  # (num_patches,)
+            "spatial_shapes": spatial
+        }
+        
+    def get_siglip_features(self, tensors_0_1):
+        dtype = torch.bfloat16
+        device = self.vae.device
+        
+        tensors_0_1 = torch.clamp(tensors_0_1, 0.0, 1.0)
+
+        mean = torch.tensor(self.image_processor.image_mean).to(
+            device, dtype=dtype
+        ).detach()
+        std = torch.tensor(self.image_processor.image_std).to(
+            device, dtype=dtype
+        ).detach()
+        # tensors_0_1 = torch.clip((255. * tensors_0_1), 0, 255).round() / 255.0
+        clip_image = (tensors_0_1 - mean.view([1, 3, 1, 1])) / std.view([1, 3, 1, 1])
+        
+        encoder_kwargs = self.tensors_to_siglip_like_features(clip_image)
+        id_embeds = self.vision_encoder(
+            pixel_values=encoder_kwargs['pixel_values'],
+            pixel_attention_mask=encoder_kwargs['pixel_attention_mask'],
+            spatial_shapes=encoder_kwargs['spatial_shapes'],
+            output_hidden_states=True,
+        )
+
+        # embeds = id_embeds['hidden_states'][-2]  # penultimate layer
+        embeds = id_embeds['pooler_output']
+        return embeds
+
+    def forward(
+        self,
+        noise,
+        noise_pred,
+        noisy_latents,
+        timesteps, 
+        batch: DataLoaderBatchDTO, 
+        scheduler: CustomFlowMatchEulerDiscreteScheduler,
+        clip_weight=1.0,
+        mse_weight=0.0,
+        model=None
+    ):
+        dtype = torch.bfloat16
+        device = self.vae.device
+        tensors = batch.tensor.to(device, dtype=dtype)
+        is_video = False
+        # stack time for video models on the batch dimension
+        if len(noise_pred.shape) == 5:
+            # B, C, T, H, W = images.shape
+            # only take first time
+            noise = noise[:, :, 0, :, :]
+            noise_pred = noise_pred[:, :, 0, :, :]
+            noisy_latents = noisy_latents[:, :, 0, :, :]
+            is_video = True
+        
+        if len(tensors.shape) == 5:
+            # batch is different
+            # (B, T, C, H, W)
+            # only take first time
+            tensors = tensors[:, 0, :, :, :]
+            
+        if model is not None and hasattr(model, 'get_stepped_pred'):
+            stepped_latents = model.get_stepped_pred(noise_pred, noise)
+        else:
+            # stepped_latents = noise - noise_pred
+            # first we step the scheduler from current timestep to the very end for a full denoise
+            bs = noise_pred.shape[0]
+            noise_pred_chunks = torch.chunk(noise_pred, bs)
+            timestep_chunks = torch.chunk(timesteps, bs)
+            noisy_latent_chunks = torch.chunk(noisy_latents, bs)
+            stepped_chunks = []
+            for idx in range(bs):
+                model_output = noise_pred_chunks[idx]
+                timestep = timestep_chunks[idx]
+                scheduler._step_index = None
+                scheduler._init_step_index(timestep)
+                sample = noisy_latent_chunks[idx].to(torch.float32)
+                
+                sigma = scheduler.sigmas[scheduler.step_index]
+                sigma_next = scheduler.sigmas[-1] # use last sigma for final step
+                prev_sample = sample + (sigma_next - sigma) * model_output
+                stepped_chunks.append(prev_sample)
+            
+            stepped_latents = torch.cat(stepped_chunks, dim=0)
+            
+        latents = stepped_latents.to(self.vae.device, dtype=self.vae.dtype)
+        
+        scaling_factor = self.vae.config['scaling_factor'] if 'scaling_factor' in self.vae.config else 1.0
+        shift_factor = self.vae.config['shift_factor'] if 'shift_factor' in self.vae.config else 0.0
+        latents = (latents / scaling_factor) + shift_factor
+        if is_video:
+            # if video, we need to unsqueeze the latents to match the vae input shape
+            latents = latents.unsqueeze(2)
+        tensors_n1p1 = self.vae.decode(latents).sample  # -1 to 1
+        
+        if is_video:
+            # if video, we need to squeeze the tensors to match the output shape
+            tensors_n1p1 = tensors_n1p1.squeeze(2)
+        
+        pred_images = (tensors_n1p1 + 1) / 2  # 0 to 1
+        
+        total_loss = 0
+        
+        with torch.no_grad():
+            target_img = tensors.to(device, dtype=dtype)
+            # go from -1 to 1 to 0 to 1
+            target_img = (target_img + 1) / 2
+            if clip_weight > 0:
+                target_clip_output = self.get_siglip_features(target_img).detach()
+        if clip_weight > 0:
+            pred_clip_output = self.get_siglip_features(pred_images)
+            clip_loss = torch.nn.functional.mse_loss(
+                pred_clip_output.float(), target_clip_output.float()
+            ) * clip_weight
+            
+            if 'clip_loss' not in self.losses:
+                self.losses['clip_loss'] = clip_loss.item()
+            else:
+                self.losses['clip_loss'] += clip_loss.item()
+            
+            total_loss += clip_loss
+        if mse_weight > 0:
+            mse_loss = torch.nn.functional.mse_loss(
+                pred_images.float(), target_img.float()
+            ) * mse_weight
+            
+            if 'mse_loss' not in self.losses:
+                self.losses['mse_loss'] = mse_loss.item()
+            else:
+                self.losses['mse_loss'] += mse_loss.item()
+            
+            total_loss += mse_loss
+
+        if self.step % self.log_every == 0 and self.step > 0:
+            print(f"DFE losses:")
+            for key in self.losses:
+                self.losses[key] /= self.log_every
+                # print in 2.000e-01 format
+                print(f" - {key}: {self.losses[key]:.3e}")
+            self.losses[key] = 0.0
+        
+        # total_loss += mse_loss
+        self.step += 1
+        
+        return total_loss
 
 def load_dfe(model_path, vae=None) -> DiffusionFeatureExtractor:
     if model_path == "v3":
         dfe = DiffusionFeatureExtractor3(vae=vae)
         dfe.eval()
         return dfe
+    if model_path == "v4":
+        dfe = DiffusionFeatureExtractor4(vae=vae)
+        dfe.eval()
+        return dfe
     if not os.path.exists(model_path):
         raise FileNotFoundError(f"Model file not found: {model_path}")
     # if it ende with safetensors

toolkit/models/wan21/autoencoder_kl_wan.py

@@ -0,0 +1,865 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import logging
+from diffusers.utils.accelerate_utils import apply_forward_hook
+from diffusers.models.activations import get_activation
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution
+import copy
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+CACHE_T = 2
+
+
+class WanCausalConv3d(nn.Conv3d):
+    r"""
+    A custom 3D causal convolution layer with feature caching support.
+
+    This layer extends the standard Conv3D layer by ensuring causality in the time dimension and handling feature
+    caching for efficient inference.
+
+    Args:
+        in_channels (int): Number of channels in the input image
+        out_channels (int): Number of channels produced by the convolution
+        kernel_size (int or tuple): Size of the convolving kernel
+        stride (int or tuple, optional): Stride of the convolution. Default: 1
+        padding (int or tuple, optional): Zero-padding added to all three sides of the input. Default: 0
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        padding: Union[int, Tuple[int, int, int]] = 0,
+    ) -> None:
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+        )
+
+        # Set up causal padding
+        self._padding = (self.padding[2], self.padding[2], self.padding[1], self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+        return super().forward(x)
+
+
+class WanRMS_norm(nn.Module):
+    r"""
+    A custom RMS normalization layer.
+
+    Args:
+        dim (int): The number of dimensions to normalize over.
+        channel_first (bool, optional): Whether the input tensor has channels as the first dimension.
+            Default is True.
+        images (bool, optional): Whether the input represents image data. Default is True.
+        bias (bool, optional): Whether to include a learnable bias term. Default is False.
+    """
+
+    def __init__(self, dim: int, channel_first: bool = True, images: bool = True, bias: bool = False) -> None:
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.0
+
+    def forward(self, x):
+        return F.normalize(x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma + self.bias
+
+
+class WanUpsample(nn.Upsample):
+    r"""
+    Perform upsampling while ensuring the output tensor has the same data type as the input.
+
+    Args:
+        x (torch.Tensor): Input tensor to be upsampled.
+
+    Returns:
+        torch.Tensor: Upsampled tensor with the same data type as the input.
+    """
+
+    def forward(self, x):
+        return super().forward(x.float()).type_as(x)
+
+
+class WanResample(nn.Module):
+    r"""
+    A custom resampling module for 2D and 3D data.
+
+    Args:
+        dim (int): The number of input/output channels.
+        mode (str): The resampling mode. Must be one of:
+            - 'none': No resampling (identity operation).
+            - 'upsample2d': 2D upsampling with nearest-exact interpolation and convolution.
+            - 'upsample3d': 3D upsampling with nearest-exact interpolation, convolution, and causal 3D convolution.
+            - 'downsample2d': 2D downsampling with zero-padding and convolution.
+            - 'downsample3d': 3D downsampling with zero-padding, convolution, and causal 3D convolution.
+    """
+
+    def __init__(self, dim: int, mode: str) -> None:
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                WanUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"), nn.Conv2d(dim, dim // 2, 3, padding=1)
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                WanUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"), nn.Conv2d(dim, dim // 2, 3, padding=1)
+            )
+            self.time_conv = WanCausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = WanCausalConv3d(dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] != "Rep":
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2
+                        )
+                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] == "Rep":
+                        cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device), cache_x], dim=2)
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]), 3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.resample(x)
+        x = x.view(b, t, x.size(1), x.size(2), x.size(3)).permute(0, 2, 1, 3, 4)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class WanResidualBlock(nn.Module):
+    r"""
+    A custom residual block module.
+
+    Args:
+        in_dim (int): Number of input channels.
+        out_dim (int): Number of output channels.
+        dropout (float, optional): Dropout rate for the dropout layer. Default is 0.0.
+        non_linearity (str, optional): Type of non-linearity to use. Default is "silu".
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        dropout: float = 0.0,
+        non_linearity: str = "silu",
+    ) -> None:
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.nonlinearity = get_activation(non_linearity)
+
+        # layers
+        self.norm1 = WanRMS_norm(in_dim, images=False)
+        self.conv1 = WanCausalConv3d(in_dim, out_dim, 3, padding=1)
+        self.norm2 = WanRMS_norm(out_dim, images=False)
+        self.dropout = nn.Dropout(dropout)
+        self.conv2 = WanCausalConv3d(out_dim, out_dim, 3, padding=1)
+        self.conv_shortcut = WanCausalConv3d(in_dim, out_dim, 1) if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        # Apply shortcut connection
+        h = self.conv_shortcut(x)
+
+        # First normalization and activation
+        x = self.norm1(x)
+        x = self.nonlinearity(x)
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        # Second normalization and activation
+        x = self.norm2(x)
+        x = self.nonlinearity(x)
+
+        # Dropout
+        x = self.dropout(x)
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+            x = self.conv2(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv2(x)
+
+        # Add residual connection
+        return x + h
+
+
+class WanAttentionBlock(nn.Module):
+    r"""
+    Causal self-attention with a single head.
+
+    Args:
+        dim (int): The number of channels in the input tensor.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = WanRMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+    def forward(self, x):
+        identity = x
+        batch_size, channels, time, height, width = x.size()
+
+        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * time, channels, height, width)
+        x = self.norm(x)
+
+        # compute query, key, value
+        qkv = self.to_qkv(x)
+        qkv = qkv.reshape(batch_size * time, 1, channels * 3, -1)
+        qkv = qkv.permute(0, 1, 3, 2).contiguous()
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(q, k, v)
+
+        x = x.squeeze(1).permute(0, 2, 1).reshape(batch_size * time, channels, height, width)
+
+        # output projection
+        x = self.proj(x)
+
+        # Reshape back: [(b*t), c, h, w] -> [b, c, t, h, w]
+        x = x.view(batch_size, time, channels, height, width)
+        x = x.permute(0, 2, 1, 3, 4)
+
+        return x + identity
+
+
+class WanMidBlock(nn.Module):
+    """
+    Middle block for WanVAE encoder and decoder.
+
+    Args:
+        dim (int): Number of input/output channels.
+        dropout (float): Dropout rate.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(self, dim: int, dropout: float = 0.0, non_linearity: str = "silu", num_layers: int = 1):
+        super().__init__()
+        self.dim = dim
+
+        # Create the components
+        resnets = [WanResidualBlock(dim, dim, dropout, non_linearity)]
+        attentions = []
+        for _ in range(num_layers):
+            attentions.append(WanAttentionBlock(dim))
+            resnets.append(WanResidualBlock(dim, dim, dropout, non_linearity))
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        # First residual block
+        x = self.resnets[0](x, feat_cache, feat_idx)
+
+        # Process through attention and residual blocks
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                x = attn(x)
+
+            x = resnet(x, feat_cache, feat_idx)
+
+        return x
+
+
+class WanEncoder3d(nn.Module):
+    r"""
+    A 3D encoder module.
+
+    Args:
+        dim (int): The base number of channels in the first layer.
+        z_dim (int): The dimensionality of the latent space.
+        dim_mult (list of int): Multipliers for the number of channels in each block.
+        num_res_blocks (int): Number of residual blocks in each block.
+        attn_scales (list of float): Scales at which to apply attention mechanisms.
+        temperal_downsample (list of bool): Whether to downsample temporally in each block.
+        dropout (float): Dropout rate for the dropout layers.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.nonlinearity = get_activation(non_linearity)
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv_in = WanCausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        self.down_blocks = nn.ModuleList([])
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                self.down_blocks.append(WanResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    self.down_blocks.append(WanAttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = "downsample3d" if temperal_downsample[i] else "downsample2d"
+                self.down_blocks.append(WanResample(out_dim, mode=mode))
+                scale /= 2.0
+
+        # middle blocks
+        self.mid_block = WanMidBlock(out_dim, dropout, non_linearity, num_layers=1)
+
+        # output blocks
+        self.norm_out = WanRMS_norm(out_dim, images=False)
+        self.conv_out = WanCausalConv3d(out_dim, z_dim, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_in(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_in(x)
+
+        ## downsamples
+        for layer in self.down_blocks:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        x = self.mid_block(x, feat_cache, feat_idx)
+
+        ## head
+        x = self.norm_out(x)
+        x = self.nonlinearity(x)
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_out(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_out(x)
+        return x
+
+
+class WanUpBlock(nn.Module):
+    """
+    A block that handles upsampling for the WanVAE decoder.
+
+    Args:
+        in_dim (int): Input dimension
+        out_dim (int): Output dimension
+        num_res_blocks (int): Number of residual blocks
+        dropout (float): Dropout rate
+        upsample_mode (str, optional): Mode for upsampling ('upsample2d' or 'upsample3d')
+        non_linearity (str): Type of non-linearity to use
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        num_res_blocks: int,
+        dropout: float = 0.0,
+        upsample_mode: Optional[str] = None,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # Create layers list
+        resnets = []
+        # Add residual blocks and attention if needed
+        current_dim = in_dim
+        for _ in range(num_res_blocks + 1):
+            resnets.append(WanResidualBlock(current_dim, out_dim, dropout, non_linearity))
+            current_dim = out_dim
+
+        self.resnets = nn.ModuleList(resnets)
+
+        # Add upsampling layer if needed
+        self.upsamplers = None
+        if upsample_mode is not None:
+            self.upsamplers = nn.ModuleList([WanResample(out_dim, mode=upsample_mode)])
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        """
+        Forward pass through the upsampling block.
+
+        Args:
+            x (torch.Tensor): Input tensor
+            feat_cache (list, optional): Feature cache for causal convolutions
+            feat_idx (list, optional): Feature index for cache management
+
+        Returns:
+            torch.Tensor: Output tensor
+        """
+        for resnet in self.resnets:
+            if feat_cache is not None:
+                x = resnet(x, feat_cache, feat_idx)
+            else:
+                x = resnet(x)
+
+        if self.upsamplers is not None:
+            if feat_cache is not None:
+                x = self.upsamplers[0](x, feat_cache, feat_idx)
+            else:
+                x = self.upsamplers[0](x)
+        return x
+
+
+class WanDecoder3d(nn.Module):
+    r"""
+    A 3D decoder module.
+
+    Args:
+        dim (int): The base number of channels in the first layer.
+        z_dim (int): The dimensionality of the latent space.
+        dim_mult (list of int): Multipliers for the number of channels in each block.
+        num_res_blocks (int): Number of residual blocks in each block.
+        attn_scales (list of float): Scales at which to apply attention mechanisms.
+        temperal_upsample (list of bool): Whether to upsample temporally in each block.
+        dropout (float): Dropout rate for the dropout layers.
+        non_linearity (str): Type of non-linearity to use.
+    """
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+        non_linearity: str = "silu",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+
+        # init block
+        self.conv_in = WanCausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.mid_block = WanMidBlock(dims[0], dropout, non_linearity, num_layers=1)
+
+        # upsample blocks
+        self.up_blocks = nn.ModuleList([])
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i > 0:
+                in_dim = in_dim // 2
+
+            # Determine if we need upsampling
+            upsample_mode = None
+            if i != len(dim_mult) - 1:
+                upsample_mode = "upsample3d" if temperal_upsample[i] else "upsample2d"
+
+            # Create and add the upsampling block
+            up_block = WanUpBlock(
+                in_dim=in_dim,
+                out_dim=out_dim,
+                num_res_blocks=num_res_blocks,
+                dropout=dropout,
+                upsample_mode=upsample_mode,
+                non_linearity=non_linearity,
+            )
+            self.up_blocks.append(up_block)
+
+            # Update scale for next iteration
+            if upsample_mode is not None:
+                scale *= 2.0
+
+        # output blocks
+        self.norm_out = WanRMS_norm(out_dim, images=False)
+        self.conv_out = WanCausalConv3d(out_dim, 3, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_in(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_in(x)
+
+        ## middle
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            # middle
+            x = self._gradient_checkpointing_func(self.mid_block, x, feat_cache, feat_idx)
+            
+            ## upsamples
+            for up_block in self.up_blocks:
+                x = self._gradient_checkpointing_func(up_block, x, feat_cache, feat_idx)
+                
+        else:
+            x = self.mid_block(x, feat_cache, feat_idx)
+
+            ## upsamples
+            for up_block in self.up_blocks:
+                x = up_block(x, feat_cache, feat_idx)
+
+        ## head
+        x = self.norm_out(x)
+        x = self.nonlinearity(x)
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv_out(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv_out(x)
+        return x
+
+
+class AutoencoderKLWan(ModelMixin, ConfigMixin):
+    r"""
+    A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos.
+    Introduced in [Wan 2.1].
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        base_dim: int = 96,
+        z_dim: int = 16,
+        dim_mult: Tuple[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: List[float] = [],
+        temperal_downsample: List[bool] = [False, True, True],
+        dropout: float = 0.0,
+        latents_mean: List[float] = [
+            -0.7571,
+            -0.7089,
+            -0.9113,
+            0.1075,
+            -0.1745,
+            0.9653,
+            -0.1517,
+            1.5508,
+            0.4134,
+            -0.0715,
+            0.5517,
+            -0.3632,
+            -0.1922,
+            -0.9497,
+            0.2503,
+            -0.2921,
+        ],
+        latents_std: List[float] = [
+            2.8184,
+            1.4541,
+            2.3275,
+            2.6558,
+            1.2196,
+            1.7708,
+            2.6052,
+            2.0743,
+            3.2687,
+            2.1526,
+            2.8652,
+            1.5579,
+            1.6382,
+            1.1253,
+            2.8251,
+            1.9160,
+        ],
+    ) -> None:
+        super().__init__()
+
+        self.z_dim = z_dim
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        self.encoder = WanEncoder3d(
+            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout
+        )
+        self.quant_conv = WanCausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.post_quant_conv = WanCausalConv3d(z_dim, z_dim, 1)
+
+        self.decoder = WanDecoder3d(
+            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout
+        )
+
+    def clear_cache(self):
+        def _count_conv3d(model):
+            count = 0
+            for m in model.modules():
+                if isinstance(m, WanCausalConv3d):
+                    count += 1
+            return count
+
+        self._conv_num = _count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = _count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+    def _encode(self, x: torch.Tensor) -> torch.Tensor:
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :], feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1) : 1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+
+        enc = self.quant_conv(out)
+        mu, logvar = enc[:, : self.z_dim, :, :, :], enc[:, self.z_dim :, :, :, :]
+        enc = torch.cat([mu, logvar], dim=1)
+        self.clear_cache()
+        return enc
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        r"""
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded videos. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        h = self._encode(x)
+        posterior = DiagonalGaussianDistribution(h)
+        if not return_dict:
+            return (posterior,)
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        self.clear_cache()
+
+        iter_ = z.shape[2]
+        x = self.post_quant_conv(z)
+        for i in range(iter_):
+            
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+
+        out = torch.clamp(out, min=-1.0, max=1.0)
+        self.clear_cache()
+        if not return_dict:
+            return (out,)
+
+        return DecoderOutput(sample=out)
+
+    @apply_forward_hook
+    def decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        decoded = self._decode(z).sample
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        """
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, return_dict=return_dict)
+        return dec

toolkit/models/wan21/wan21.py

@@ -9,7 +9,8 @@ from toolkit.dequantize import patch_dequantization_on_save
 from toolkit.models.base_model import BaseModel
 from toolkit.prompt_utils import PromptEmbeds
 from transformers import AutoTokenizer, UMT5EncoderModel
-from diffusers import AutoencoderKLWan, WanPipeline, WanTransformer3DModel
+from diffusers import  WanPipeline, WanTransformer3DModel, AutoencoderKL
+from .autoencoder_kl_wan import AutoencoderKLWan
 import os
 import sys