Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+test_image.jpg filter=lfs diff=lfs merge=lfs -text

model_index.json

@@ -0,0 +1,16 @@
+{
+  "_class_name": "PixCellPipeline",
+  "_diffusers_version": "0.32.2",
+  "transformer": [
+    "pixcell_transformer_2d",
+    "PixCellTransformer2DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKL"
+  ],
+  "scheduler": [
+    "diffusers",
+    "DPMSolverMultistepScheduler"
+  ]
+}

pipeline.py

@@ -0,0 +1,1122 @@
+# Copyright 2024 PixArt-Sigma Authors 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.
+
+import inspect
+import urllib.parse as ul
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from diffusers.image_processor import PixArtImageProcessor
+from diffusers.models import AutoencoderKL
+from diffusers.schedulers import DPMSolverMultistepScheduler
+from diffusers.utils import (
+    BACKENDS_MAPPING,
+    deprecate,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+from pixcell_transformer_2d import PixCellTransformer2DModel
+
+
+from typing import Any, Dict, Optional, Union
+
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import is_torch_version, logging
+from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.models.activations import deprecate, FP32SiLU
+
+
+def pixcell_get_2d_sincos_pos_embed(
+    embed_dim,
+    grid_size,
+    cls_token=False,
+    extra_tokens=0,
+    interpolation_scale=1.0,
+    base_size=16,
+    device: Optional[torch.device] = None,
+    phase=0,
+    output_type: str = "np",
+):
+    """
+    Creates 2D sinusoidal positional embeddings.
+
+    Args:
+        embed_dim (`int`):
+            The embedding dimension.
+        grid_size (`int`):
+            The size of the grid height and width.
+        cls_token (`bool`, defaults to `False`):
+            Whether or not to add a classification token.
+        extra_tokens (`int`, defaults to `0`):
+            The number of extra tokens to add.
+        interpolation_scale (`float`, defaults to `1.0`):
+            The scale of the interpolation.
+
+    Returns:
+        pos_embed (`torch.Tensor`):
+            Shape is either `[grid_size * grid_size, embed_dim]` if not using cls_token, or `[1 + grid_size*grid_size,
+            embed_dim]` if using cls_token
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if isinstance(grid_size, int):
+        grid_size = (grid_size, grid_size)
+
+    grid_h = (
+        torch.arange(grid_size[0], device=device, dtype=torch.float32)
+        / (grid_size[0] / base_size)
+        / interpolation_scale
+    )
+    grid_w = (
+        torch.arange(grid_size[1], device=device, dtype=torch.float32)
+        / (grid_size[1] / base_size)
+        / interpolation_scale
+    )
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")  # here w goes first
+    grid = torch.stack(grid, dim=0)
+
+    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
+    pos_embed = pixcell_get_2d_sincos_pos_embed_from_grid(embed_dim, grid, phase=phase, output_type=output_type)
+    if cls_token and extra_tokens > 0:
+        pos_embed = torch.concat([torch.zeros([extra_tokens, embed_dim]), pos_embed], dim=0)
+    return pos_embed
+
+
+def pixcell_get_2d_sincos_pos_embed_from_grid(embed_dim, grid, phase=0, output_type="np"):
+    r"""
+    This function generates 2D sinusoidal positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension.
+        grid (`torch.Tensor`): Grid of positions with shape `(H * W,)`.
+
+    Returns:
+        `torch.Tensor`: The 2D sinusoidal positional embeddings with shape `(H * W, embed_dim)`
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    # use half of dimensions to encode grid_h
+    emb_h = pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0], phase=phase, output_type=output_type)  # (H*W, D/2)
+    emb_w = pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1], phase=phase, output_type=output_type)  # (H*W, D/2)
+
+    emb = torch.concat([emb_h, emb_w], dim=1)  # (H*W, D)
+    return emb
+
+
+def pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim, pos, phase=0, output_type="np"):
+    """
+    This function generates 1D positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension `D`
+        pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)`
+
+    Returns:
+        `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`.
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_1d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.34.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1) + phase  # (M,)
+    out = torch.outer(pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
+
+
+class PixcellUNIProjection(nn.Module):
+    """
+    Projects UNI embeddings. Also handles dropout for classifier-free guidance.
+
+    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+    """
+
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", num_tokens=1):
+        super().__init__()
+        if out_features is None:
+            out_features = hidden_size
+        self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
+        if act_fn == "gelu_tanh":
+            self.act_1 = nn.GELU(approximate="tanh")
+        elif act_fn == "silu":
+            self.act_1 = nn.SiLU()
+        elif act_fn == "silu_fp32":
+            self.act_1 = FP32SiLU()
+        else:
+            raise ValueError(f"Unknown activation function: {act_fn}")
+        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True)
+
+        self.register_buffer("uncond_embedding", nn.Parameter(torch.randn(num_tokens, in_features) / in_features ** 0.5))
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+class UNIPosEmbed(nn.Module):
+    """
+    Adds positional embeddings to the UNI conditions.
+
+    Args:
+        height (`int`, defaults to `224`): The height of the image.
+        width (`int`, defaults to `224`): The width of the image.
+        patch_size (`int`, defaults to `16`): The size of the patches.
+        in_channels (`int`, defaults to `3`): The number of input channels.
+        embed_dim (`int`, defaults to `768`): The output dimension of the embedding.
+        layer_norm (`bool`, defaults to `False`): Whether or not to use layer normalization.
+        flatten (`bool`, defaults to `True`): Whether or not to flatten the output.
+        bias (`bool`, defaults to `True`): Whether or not to use bias.
+        interpolation_scale (`float`, defaults to `1`): The scale of the interpolation.
+        pos_embed_type (`str`, defaults to `"sincos"`): The type of positional embedding.
+        pos_embed_max_size (`int`, defaults to `None`): The maximum size of the positional embedding.
+    """
+
+    def __init__(
+        self,
+        height=1,
+        width=1,
+        base_size=16,
+        embed_dim=768,
+        interpolation_scale=1,
+        pos_embed_type="sincos",
+    ):
+        super().__init__()
+
+        num_embeds = height*width
+        grid_size = int(num_embeds ** 0.5)
+
+        if pos_embed_type == "sincos":
+            y_pos_embed = pixcell_get_2d_sincos_pos_embed(
+                embed_dim,
+                grid_size,
+                base_size=base_size,
+                interpolation_scale=interpolation_scale,
+                output_type="pt",
+                phase = base_size // num_embeds
+            )
+            self.register_buffer("y_pos_embed", y_pos_embed.float().unsqueeze(0))
+        else:
+            raise ValueError("`pos_embed_type` not supported")
+
+    def forward(self, uni_embeds):
+        return (uni_embeds + self.y_pos_embed).to(uni_embeds.dtype)
+
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class PixCellTransformer2DModel(ModelMixin, ConfigMixin):
+    r"""
+    A 2D Transformer model as introduced in PixArt family of models (https://arxiv.org/abs/2310.00426,
+    https://arxiv.org/abs/2403.04692). Modified for the pathology domain.
+
+    Parameters:
+        num_attention_heads (int, optional, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (int, optional, defaults to 72): The number of channels in each head.
+        in_channels (int, defaults to 4): The number of channels in the input.
+        out_channels (int, optional):
+            The number of channels in the output. Specify this parameter if the output channel number differs from the
+            input.
+        num_layers (int, optional, defaults to 28): The number of layers of Transformer blocks to use.
+        dropout (float, optional, defaults to 0.0): The dropout probability to use within the Transformer blocks.
+        norm_num_groups (int, optional, defaults to 32):
+            Number of groups for group normalization within Transformer blocks.
+        cross_attention_dim (int, optional):
+            The dimensionality for cross-attention layers, typically matching the encoder's hidden dimension.
+        attention_bias (bool, optional, defaults to True):
+            Configure if the Transformer blocks' attention should contain a bias parameter.
+        sample_size (int, defaults to 128):
+            The width of the latent images. This parameter is fixed during training.
+        patch_size (int, defaults to 2):
+            Size of the patches the model processes, relevant for architectures working on non-sequential data.
+        activation_fn (str, optional, defaults to "gelu-approximate"):
+            Activation function to use in feed-forward networks within Transformer blocks.
+        num_embeds_ada_norm (int, optional, defaults to 1000):
+            Number of embeddings for AdaLayerNorm, fixed during training and affects the maximum denoising steps during
+            inference.
+        upcast_attention (bool, optional, defaults to False):
+            If true, upcasts the attention mechanism dimensions for potentially improved performance.
+        norm_type (str, optional, defaults to "ada_norm_zero"):
+            Specifies the type of normalization used, can be 'ada_norm_zero'.
+        norm_elementwise_affine (bool, optional, defaults to False):
+            If true, enables element-wise affine parameters in the normalization layers.
+        norm_eps (float, optional, defaults to 1e-6):
+            A small constant added to the denominator in normalization layers to prevent division by zero.
+        interpolation_scale (int, optional): Scale factor to use during interpolating the position embeddings.
+        use_additional_conditions (bool, optional): If we're using additional conditions as inputs.
+        attention_type (str, optional, defaults to "default"): Kind of attention mechanism to be used.
+        caption_channels (int, optional, defaults to None):
+            Number of channels to use for projecting the caption embeddings.
+        use_linear_projection (bool, optional, defaults to False):
+            Deprecated argument. Will be removed in a future version.
+        num_vector_embeds (bool, optional, defaults to False):
+            Deprecated argument. Will be removed in a future version.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "PatchEmbed"]
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 72,
+        in_channels: int = 4,
+        out_channels: Optional[int] = 8,
+        num_layers: int = 28,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = 1152,
+        attention_bias: bool = True,
+        sample_size: int = 128,
+        patch_size: int = 2,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: Optional[int] = 1000,
+        upcast_attention: bool = False,
+        norm_type: str = "ada_norm_single",
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+        use_additional_conditions: Optional[bool] = None,
+        caption_channels: Optional[int] = None,
+        caption_num_tokens: int = 1,
+        attention_type: Optional[str] = "default",
+    ):
+        super().__init__()
+
+        # Validate inputs.
+        if norm_type != "ada_norm_single":
+            raise NotImplementedError(
+                f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+            )
+        elif norm_type == "ada_norm_single" and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+            )
+
+        # Set some common variables used across the board.
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if use_additional_conditions is None:
+            if sample_size == 128:
+                use_additional_conditions = True
+            else:
+                use_additional_conditions = False
+        self.use_additional_conditions = use_additional_conditions
+
+        self.gradient_checkpointing = False
+
+        # 2. Initialize the position embedding and transformer blocks.
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+
+        interpolation_scale = (
+            self.config.interpolation_scale
+            if self.config.interpolation_scale is not None
+            else max(self.config.sample_size // 64, 1)
+        )
+        self.pos_embed = PatchEmbed(
+            height=self.config.sample_size,
+            width=self.config.sample_size,
+            patch_size=self.config.patch_size,
+            in_channels=self.config.in_channels,
+            embed_dim=self.inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        # Initialize the positional embedding for the conditions for >1 UNI embeddings 
+        if self.config.caption_num_tokens == 1:
+            self.y_pos_embed = None
+        else:
+            # 1:1 aspect ratio
+            self.uni_height = int(self.config.caption_num_tokens ** 0.5)
+            self.uni_width = int(self.config.caption_num_tokens ** 0.5)
+
+            self.y_pos_embed = UNIPosEmbed(
+                height=self.uni_height,
+                width=self.uni_width,
+                base_size=self.config.sample_size // self.config.patch_size,
+                embed_dim=self.config.caption_channels,
+                interpolation_scale=2, # Should this be fixed?
+                pos_embed_type="sincos", # This is fixed
+            )
+
+        # 3. Output blocks.
+        self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim**0.5)
+        self.proj_out = nn.Linear(self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels)
+
+        self.adaln_single = AdaLayerNormSingle(
+            self.inner_dim, use_additional_conditions=self.use_additional_conditions
+        )
+        self.caption_projection = None
+        if self.config.caption_channels is not None:
+            self.caption_projection = PixcellUNIProjection(
+                in_features=self.config.caption_channels, hidden_size=self.inner_dim, num_tokens=self.config.caption_num_tokens,
+            )
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+
+        Safe to just use `AttnProcessor()` as PixArt doesn't have any exotic attention processors in default model.
+        """
+        self.set_attn_processor(AttnProcessor())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`PixCellTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep (`torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            added_cond_kwargs: (`Dict[str, Any]`, *optional*): Additional conditions to be used as inputs.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if self.use_additional_conditions and added_cond_kwargs is None:
+            raise ValueError("`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`.")
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size = hidden_states.shape[0]
+        height, width = (
+            hidden_states.shape[-2] // self.config.patch_size,
+            hidden_states.shape[-1] // self.config.patch_size,
+        )
+        hidden_states = self.pos_embed(hidden_states)
+
+        timestep, embedded_timestep = self.adaln_single(
+            timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        if self.caption_projection is not None:
+            # Add positional embeddings to conditions if >1 UNI are given
+            if self.y_pos_embed is not None:
+                encoder_hidden_states = self.y_pos_embed(encoder_hidden_states)
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    None,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=None,
+                )
+
+        # 3. Output
+        shift, scale = (
+            self.scale_shift_table[None] + embedded_timestep[:, None].to(self.scale_shift_table.device)
+        ).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+        # Modulation
+        hidden_states = hidden_states * (1 + scale.to(hidden_states.device)) + shift.to(hidden_states.device)
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        hidden_states = hidden_states.reshape(
+            shape=(-1, height, width, self.config.patch_size, self.config.patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(-1, self.out_channels, height * self.config.patch_size, width * self.config.patch_size)
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
+
+
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import PixCellSigmaPipeline
+
+        >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-Sigma-XL-2-512-MS" too.
+        >>> pipe = PixArtSigmaPipeline.from_pretrained(
+        ...     "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS", torch_dtype=torch.float16
+        ... )
+        >>> # Enable memory optimizations.
+        >>> # pipe.enable_model_cpu_offload()
+
+        >>> prompt = "A small cactus with a happy face in the Sahara desert."
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class PixCellPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for SSL-to-image generation using PixCell.
+    """
+
+    model_cpu_offload_seq = "transformer->vae"
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        transformer: PixCellTransformer2DModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def get_unconditional_embedding(self, batch_size=1):
+        # Unconditional embedding is learned
+        uncond = self.transformer.caption_projection.uncond_embedding.clone().tile(batch_size,1,1)
+        return uncond
+
+    # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline.check_inputs
+    def check_inputs(
+        self,
+        height,
+        width,
+        callback_steps,
+        uni_embeds=None,
+        negative_uni_embeds=None,
+        guidance_scale=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if uni_embeds is None:
+            raise ValueError(
+                "Provide a UNI embedding `uni_embeds`."
+            )
+        elif len(uni_embeds.shape) != 3:
+            raise ValueError(
+                "UNI embedding given is not in (B,N,D)."
+            )
+        elif uni_embeds.shape[1] != self.transformer.config.caption_num_tokens:
+            raise ValueError(
+                f"Number of UNI embeddings must match the ones used in training ({self.transformer.config.caption_num_tokens})."
+            )
+        elif uni_embeds.shape[2] != self.transformer.config.caption_channels:
+            raise ValueError(
+                "UNI embedding given has incorrect dimenions."
+            )
+
+        if guidance_scale > 1.0:
+            if negative_uni_embeds is None:
+                raise ValueError(
+                    "Provide a negative UNI embedding `negative_uni_embeds`."
+                )
+            elif len(negative_uni_embeds.shape) != 3:
+                raise ValueError(
+                    "Negative UNI embedding given is not in (B,N,D)."
+                )
+            elif negative_uni_embeds.shape[1] != self.transformer.config.caption_num_tokens:
+                raise ValueError(
+                    f"Number of negative UNI embeddings must match the ones used in training ({self.transformer.config.caption_num_tokens})."
+                )
+            elif negative_uni_embeds.shape[2] != self.transformer.config.caption_channels:
+                raise ValueError(
+                    "Negative UNI embedding given has incorrect dimenions."
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        num_inference_steps: int = 20,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
+        guidance_scale: float = 1.5,
+        num_images_per_prompt: Optional[int] = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        uni_embeds: Optional[torch.Tensor] = None,
+        negative_uni_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
+        callback_steps: int = 1,
+        **kwargs,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            num_inference_steps (`int`, *optional*, defaults to 100):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 4.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            uni_embeds (`torch.Tensor`, *optional*):
+                Pre-generated UNI embeddings. 
+            negative_uni_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative UNI embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ImagePipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images
+        """
+        # 1. Check inputs. Raise error if not correct
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor
+
+        self.check_inputs(
+            height,
+            width,
+            callback_steps,
+            uni_embeds,
+            negative_uni_embeds,
+            guidance_scale,
+        )
+
+        # 2. Default height and width to transformer
+        batch_size = uni_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Handle UNI conditioning
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        uni_embeds = uni_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        if do_classifier_free_guidance:
+            negative_uni_embeds = negative_uni_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uni_embeds = torch.cat([negative_uni_embeds, uni_embeds], dim=0)
+
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas
+        )
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            uni_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        added_cond_kwargs = {}
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                current_timestep = t
+                if not torch.is_tensor(current_timestep):
+                    # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+                    # This would be a good case for the `match` statement (Python 3.10+)
+                    is_mps = latent_model_input.device.type == "mps"
+                    if isinstance(current_timestep, float):
+                        dtype = torch.float32 if is_mps else torch.float64
+                    else:
+                        dtype = torch.int32 if is_mps else torch.int64
+                    current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
+                elif len(current_timestep.shape) == 0:
+                    current_timestep = current_timestep[None].to(latent_model_input.device)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                current_timestep = current_timestep.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=uni_embeds,
+                    timestep=current_timestep,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # learned sigma
+                if self.transformer.config.out_channels // 2 == latent_channels:
+                    noise_pred = noise_pred.chunk(2, dim=1)[0]
+                else:
+                    noise_pred = noise_pred
+
+                # compute previous image: x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            vae_scale = self.vae.config.scaling_factor
+            vae_shift = getattr(self.vae.config, "shift_factor", 0)
+
+            image = self.vae.decode((latents / vae_scale) + vae_shift, return_dict=False)[0]
+
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

scheduler/scheduler_config.json

@@ -0,0 +1,30 @@
+{
+  "_class_name": "DPMSolverMultistepScheduler",
+  "_diffusers_version": "0.32.2",
+  "algorithm_type": "dpmsolver++",
+  "beta_end": 0.02,
+  "beta_schedule": "linear",
+  "beta_start": 0.0001,
+  "dynamic_thresholding_ratio": 0.995,
+  "euler_at_final": false,
+  "final_sigmas_type": "zero",
+  "flow_shift": 1.0,
+  "lambda_min_clipped": -Infinity,
+  "lower_order_final": true,
+  "num_train_timesteps": 1000,
+  "prediction_type": "epsilon",
+  "rescale_betas_zero_snr": false,
+  "sample_max_value": 1.0,
+  "solver_order": 2,
+  "solver_type": "midpoint",
+  "steps_offset": 0,
+  "thresholding": false,
+  "timestep_spacing": "linspace",
+  "trained_betas": null,
+  "use_beta_sigmas": false,
+  "use_exponential_sigmas": false,
+  "use_flow_sigmas": false,
+  "use_karras_sigmas": false,
+  "use_lu_lambdas": false,
+  "variance_type": null
+}

transformer/config.json

@@ -0,0 +1,31 @@
+{
+  "_class_name": "PixCellTransformer2DModel",
+  "_diffusers_version": "0.32.2",
+  "_name_or_path": "pixart_256/transformer",
+  "activation_fn": "gelu-approximate",
+  "attention_bias": true,
+  "attention_head_dim": 72,
+  "attention_type": "default",
+  "caption_channels": 1536,
+  "caption_num_tokens": 1,
+  "cross_attention_dim": 1152,
+  "double_self_attention": false,
+  "dropout": 0.0,
+  "in_channels": 16,
+  "interpolation_scale": 0.5,
+  "norm_elementwise_affine": false,
+  "norm_eps": 1e-06,
+  "norm_num_groups": 32,
+  "norm_type": "ada_norm_single",
+  "num_attention_heads": 16,
+  "num_embeds_ada_norm": 1000,
+  "num_layers": 28,
+  "num_vector_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 32,
+  "patch_size": 2,
+  "sample_size": 32,
+  "upcast_attention": false,
+  "use_additional_conditions": false,
+  "use_linear_projection": false
+}

transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2897bd89e90037faf71c16ba6ce87165cb7c8509cafdab6aa824c3ea827cbe8
+size 2432366184

transformer/pixcell_transformer_2d.py

@@ -0,0 +1,676 @@
+# Copyright 2024 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 Any, Dict, Optional, Union
+
+import torch
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import is_torch_version, logging
+from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.models.activations import deprecate, FP32SiLU
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# PixCell UNI conditioning
+def pixcell_get_2d_sincos_pos_embed(
+    embed_dim,
+    grid_size,
+    cls_token=False,
+    extra_tokens=0,
+    interpolation_scale=1.0,
+    base_size=16,
+    device: Optional[torch.device] = None,
+    phase=0,
+    output_type: str = "np",
+):
+    """
+    Creates 2D sinusoidal positional embeddings.
+
+    Args:
+        embed_dim (`int`):
+            The embedding dimension.
+        grid_size (`int`):
+            The size of the grid height and width.
+        cls_token (`bool`, defaults to `False`):
+            Whether or not to add a classification token.
+        extra_tokens (`int`, defaults to `0`):
+            The number of extra tokens to add.
+        interpolation_scale (`float`, defaults to `1.0`):
+            The scale of the interpolation.
+
+    Returns:
+        pos_embed (`torch.Tensor`):
+            Shape is either `[grid_size * grid_size, embed_dim]` if not using cls_token, or `[1 + grid_size*grid_size,
+            embed_dim]` if using cls_token
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if isinstance(grid_size, int):
+        grid_size = (grid_size, grid_size)
+
+    grid_h = (
+        torch.arange(grid_size[0], device=device, dtype=torch.float32)
+        / (grid_size[0] / base_size)
+        / interpolation_scale
+    )
+    grid_w = (
+        torch.arange(grid_size[1], device=device, dtype=torch.float32)
+        / (grid_size[1] / base_size)
+        / interpolation_scale
+    )
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")  # here w goes first
+    grid = torch.stack(grid, dim=0)
+
+    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
+    pos_embed = pixcell_get_2d_sincos_pos_embed_from_grid(embed_dim, grid, phase=phase, output_type=output_type)
+    if cls_token and extra_tokens > 0:
+        pos_embed = torch.concat([torch.zeros([extra_tokens, embed_dim]), pos_embed], dim=0)
+    return pos_embed
+
+
+def pixcell_get_2d_sincos_pos_embed_from_grid(embed_dim, grid, phase=0, output_type="np"):
+    r"""
+    This function generates 2D sinusoidal positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension.
+        grid (`torch.Tensor`): Grid of positions with shape `(H * W,)`.
+
+    Returns:
+        `torch.Tensor`: The 2D sinusoidal positional embeddings with shape `(H * W, embed_dim)`
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    # use half of dimensions to encode grid_h
+    emb_h = pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0], phase=phase, output_type=output_type)  # (H*W, D/2)
+    emb_w = pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1], phase=phase, output_type=output_type)  # (H*W, D/2)
+
+    emb = torch.concat([emb_h, emb_w], dim=1)  # (H*W, D)
+    return emb
+
+
+def pixcell_get_1d_sincos_pos_embed_from_grid(embed_dim, pos, phase=0, output_type="np"):
+    """
+    This function generates 1D positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension `D`
+        pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)`
+
+    Returns:
+        `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`.
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_1d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.34.0", deprecation_message, standard_warn=False)
+        raise ValueError("Not supported")
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1) + phase  # (M,)
+    out = torch.outer(pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
+
+
+class PixcellUNIProjection(nn.Module):
+    """
+    Projects UNI embeddings. Also handles dropout for classifier-free guidance.
+
+    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+    """
+
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", num_tokens=1):
+        super().__init__()
+        if out_features is None:
+            out_features = hidden_size
+        self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
+        if act_fn == "gelu_tanh":
+            self.act_1 = nn.GELU(approximate="tanh")
+        elif act_fn == "silu":
+            self.act_1 = nn.SiLU()
+        elif act_fn == "silu_fp32":
+            self.act_1 = FP32SiLU()
+        else:
+            raise ValueError(f"Unknown activation function: {act_fn}")
+        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True)
+
+        self.register_buffer("uncond_embedding", nn.Parameter(torch.randn(num_tokens, in_features) / in_features ** 0.5))
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+class UNIPosEmbed(nn.Module):
+    """
+    Adds positional embeddings to the UNI conditions.
+
+    Args:
+        height (`int`, defaults to `224`): The height of the image.
+        width (`int`, defaults to `224`): The width of the image.
+        patch_size (`int`, defaults to `16`): The size of the patches.
+        in_channels (`int`, defaults to `3`): The number of input channels.
+        embed_dim (`int`, defaults to `768`): The output dimension of the embedding.
+        layer_norm (`bool`, defaults to `False`): Whether or not to use layer normalization.
+        flatten (`bool`, defaults to `True`): Whether or not to flatten the output.
+        bias (`bool`, defaults to `True`): Whether or not to use bias.
+        interpolation_scale (`float`, defaults to `1`): The scale of the interpolation.
+        pos_embed_type (`str`, defaults to `"sincos"`): The type of positional embedding.
+        pos_embed_max_size (`int`, defaults to `None`): The maximum size of the positional embedding.
+    """
+
+    def __init__(
+        self,
+        height=1,
+        width=1,
+        base_size=16,
+        embed_dim=768,
+        interpolation_scale=1,
+        pos_embed_type="sincos",
+    ):
+        super().__init__()
+
+        num_embeds = height*width
+        grid_size = int(num_embeds ** 0.5)
+
+        if pos_embed_type == "sincos":
+            y_pos_embed = pixcell_get_2d_sincos_pos_embed(
+                embed_dim,
+                grid_size,
+                base_size=base_size,
+                interpolation_scale=interpolation_scale,
+                output_type="pt",
+                phase = base_size // num_embeds
+            )
+            self.register_buffer("y_pos_embed", y_pos_embed.float().unsqueeze(0))
+        else:
+            raise ValueError("`pos_embed_type` not supported")
+
+    def forward(self, uni_embeds):
+        return (uni_embeds + self.y_pos_embed).to(uni_embeds.dtype)
+
+
+
+class PixCellTransformer2DModel(ModelMixin, ConfigMixin):
+    r"""
+    A 2D Transformer model as introduced in PixArt family of models (https://arxiv.org/abs/2310.00426,
+    https://arxiv.org/abs/2403.04692). Modified for the pathology domain.
+
+    Parameters:
+        num_attention_heads (int, optional, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (int, optional, defaults to 72): The number of channels in each head.
+        in_channels (int, defaults to 4): The number of channels in the input.
+        out_channels (int, optional):
+            The number of channels in the output. Specify this parameter if the output channel number differs from the
+            input.
+        num_layers (int, optional, defaults to 28): The number of layers of Transformer blocks to use.
+        dropout (float, optional, defaults to 0.0): The dropout probability to use within the Transformer blocks.
+        norm_num_groups (int, optional, defaults to 32):
+            Number of groups for group normalization within Transformer blocks.
+        cross_attention_dim (int, optional):
+            The dimensionality for cross-attention layers, typically matching the encoder's hidden dimension.
+        attention_bias (bool, optional, defaults to True):
+            Configure if the Transformer blocks' attention should contain a bias parameter.
+        sample_size (int, defaults to 128):
+            The width of the latent images. This parameter is fixed during training.
+        patch_size (int, defaults to 2):
+            Size of the patches the model processes, relevant for architectures working on non-sequential data.
+        activation_fn (str, optional, defaults to "gelu-approximate"):
+            Activation function to use in feed-forward networks within Transformer blocks.
+        num_embeds_ada_norm (int, optional, defaults to 1000):
+            Number of embeddings for AdaLayerNorm, fixed during training and affects the maximum denoising steps during
+            inference.
+        upcast_attention (bool, optional, defaults to False):
+            If true, upcasts the attention mechanism dimensions for potentially improved performance.
+        norm_type (str, optional, defaults to "ada_norm_zero"):
+            Specifies the type of normalization used, can be 'ada_norm_zero'.
+        norm_elementwise_affine (bool, optional, defaults to False):
+            If true, enables element-wise affine parameters in the normalization layers.
+        norm_eps (float, optional, defaults to 1e-6):
+            A small constant added to the denominator in normalization layers to prevent division by zero.
+        interpolation_scale (int, optional): Scale factor to use during interpolating the position embeddings.
+        use_additional_conditions (bool, optional): If we're using additional conditions as inputs.
+        attention_type (str, optional, defaults to "default"): Kind of attention mechanism to be used.
+        caption_channels (int, optional, defaults to None):
+            Number of channels to use for projecting the caption embeddings.
+        use_linear_projection (bool, optional, defaults to False):
+            Deprecated argument. Will be removed in a future version.
+        num_vector_embeds (bool, optional, defaults to False):
+            Deprecated argument. Will be removed in a future version.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "PatchEmbed"]
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 72,
+        in_channels: int = 4,
+        out_channels: Optional[int] = 8,
+        num_layers: int = 28,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = 1152,
+        attention_bias: bool = True,
+        sample_size: int = 128,
+        patch_size: int = 2,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: Optional[int] = 1000,
+        upcast_attention: bool = False,
+        norm_type: str = "ada_norm_single",
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+        use_additional_conditions: Optional[bool] = None,
+        caption_channels: Optional[int] = None,
+        caption_num_tokens: int = 1,
+        attention_type: Optional[str] = "default",
+    ):
+        super().__init__()
+
+        # Validate inputs.
+        if norm_type != "ada_norm_single":
+            raise NotImplementedError(
+                f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+            )
+        elif norm_type == "ada_norm_single" and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+            )
+
+        # Set some common variables used across the board.
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if use_additional_conditions is None:
+            if sample_size == 128:
+                use_additional_conditions = True
+            else:
+                use_additional_conditions = False
+        self.use_additional_conditions = use_additional_conditions
+
+        self.gradient_checkpointing = False
+
+        # 2. Initialize the position embedding and transformer blocks.
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+
+        interpolation_scale = (
+            self.config.interpolation_scale
+            if self.config.interpolation_scale is not None
+            else max(self.config.sample_size // 64, 1)
+        )
+        self.pos_embed = PatchEmbed(
+            height=self.config.sample_size,
+            width=self.config.sample_size,
+            patch_size=self.config.patch_size,
+            in_channels=self.config.in_channels,
+            embed_dim=self.inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        # Initialize the positional embedding for the conditions for >1 UNI embeddings 
+        if self.config.caption_num_tokens == 1:
+            self.y_pos_embed = None
+        else:
+            # 1:1 aspect ratio
+            self.uni_height = int(self.config.caption_num_tokens ** 0.5)
+            self.uni_width = int(self.config.caption_num_tokens ** 0.5)
+
+            self.y_pos_embed = UNIPosEmbed(
+                height=self.uni_height,
+                width=self.uni_width,
+                base_size=self.config.sample_size // self.config.patch_size,
+                embed_dim=self.config.caption_channels,
+                interpolation_scale=2, # Should this be fixed?
+                pos_embed_type="sincos", # This is fixed
+            )
+
+        # 3. Output blocks.
+        self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim**0.5)
+        self.proj_out = nn.Linear(self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels)
+
+        self.adaln_single = AdaLayerNormSingle(
+            self.inner_dim, use_additional_conditions=self.use_additional_conditions
+        )
+        self.caption_projection = None
+        if self.config.caption_channels is not None:
+            self.caption_projection = PixcellUNIProjection(
+                in_features=self.config.caption_channels, hidden_size=self.inner_dim, num_tokens=self.config.caption_num_tokens,
+            )
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+
+        Safe to just use `AttnProcessor()` as PixArt doesn't have any exotic attention processors in default model.
+        """
+        self.set_attn_processor(AttnProcessor())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`PixCellTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep (`torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            added_cond_kwargs: (`Dict[str, Any]`, *optional*): Additional conditions to be used as inputs.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if self.use_additional_conditions and added_cond_kwargs is None:
+            raise ValueError("`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`.")
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size = hidden_states.shape[0]
+        height, width = (
+            hidden_states.shape[-2] // self.config.patch_size,
+            hidden_states.shape[-1] // self.config.patch_size,
+        )
+        hidden_states = self.pos_embed(hidden_states)
+
+        timestep, embedded_timestep = self.adaln_single(
+            timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        if self.caption_projection is not None:
+            # Add positional embeddings to conditions if >1 UNI are given
+            if self.y_pos_embed is not None:
+                encoder_hidden_states = self.y_pos_embed(encoder_hidden_states)
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    None,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=None,
+                )
+
+        # 3. Output
+        shift, scale = (
+            self.scale_shift_table[None] + embedded_timestep[:, None].to(self.scale_shift_table.device)
+        ).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+        # Modulation
+        hidden_states = hidden_states * (1 + scale.to(hidden_states.device)) + shift.to(hidden_states.device)
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        hidden_states = hidden_states.reshape(
+            shape=(-1, height, width, self.config.patch_size, self.config.patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(-1, self.out_channels, height * self.config.patch_size, width * self.config.patch_size)
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

vae/config.json

@@ -0,0 +1,38 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.32.2",
+  "_name_or_path": "stabilityai/stable-diffusion-3.5-large",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 3,
+  "latent_channels": 16,
+  "latents_mean": null,
+  "latents_std": null,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 1024,
+  "scaling_factor": 1.5305,
+  "shift_factor": 0.0609,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": false,
+  "use_quant_conv": false
+}