Upload folder using huggingface_hub

config.json

@@ -0,0 +1,23 @@
+{
+  "architectures": [
+    "HunyuanViTModel"
+  ],
+  "_attn_implementation": "flash_attention_2",
+  "initializer_factor": 1.0,
+  "model_type": "HunyuanViT",
+  "torch_dtype": "float32",
+  "transformers_version": "4.52.4",
+  "vision_config": {
+    "attention_dropout": 0.0,
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1536,
+    "intermediate_size": 6144,
+    "layer_norm_eps": 1e-06,
+    "model_type": "HunyuanViT_vision_model",
+    "num_attention_heads": 16,
+    "num_channels": 3,
+    "num_hidden_layers": 40,
+    "num_patches": 16384,
+    "patch_size": 16
+  }
+}

configuration_HunyuanViT.py

@@ -0,0 +1,179 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/HunyuanViT/modular_HunyuanViT.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_HunyuanViT.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team.
+#
+# 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 transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+class HunyuanViTVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`HunyuanViTVisionModel`]. It is used to instantiate a
+    HunyuanViT vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the HunyuanViT
+    [google/HunyuanViT-base-patch16-naflex](https://huggingface.co/google/HunyuanViT-base-patch16-naflex) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        num_patches (`int`, *optional*, defaults to 256):
+            The number of patches in the image with the size of (`patch_size`, `patch_size`).
+            The image is resized to fill maximum of this number of patches, and to preserve
+            the aspect ratio. In case the resulted number of patches is lower, the image is
+            padded in "patch" dimension.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    Example:
+
+    ```python
+    >>> from transformers import HunyuanViTVisionConfig, HunyuanViTVisionModel
+
+    >>> # Initializing a HunyuanViTVisionConfig with google/HunyuanViT-base-patch16-naflex style configuration
+    >>> configuration = HunyuanViTVisionConfig()
+
+    >>> # Initializing a HunyuanViTVisionModel (with random weights) from the google/HunyuanViT-base-patch16-naflex style configuration
+    >>> model = HunyuanViTVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "HunyuanViT_vision_model"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        num_patches=256,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.num_patches = num_patches
+
+
+class HunyuanViTConfig(PretrainedConfig):
+    r"""
+    [`HunyuanViTConfig`] is the configuration class to store the configuration of a [`HunyuanViTModel`]. It is used to
+    instantiate a HunyuanViT model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the HunyuanViT
+    [google/HunyuanViT-base-patch16-224](https://huggingface.co/google/HunyuanViT-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`HunyuanViTTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`HunyuanViTVisionConfig`].
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import HunyuanViTConfig, HunyuanViTModel
+
+    >>> # Initializing a HunyuanViTConfig with google/HunyuanViT-base-patch16-224 style configuration
+    >>> configuration = HunyuanViTConfig()
+
+    >>> # Initializing a HunyuanViTModel (with random weights) from the google/HunyuanViT-base-patch16-224 style configuration
+    >>> model = HunyuanViTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a HunyuanViTConfig from a HunyuanViTTextConfig and a HunyuanViTVisionConfig
+    >>> from transformers import HunyuanViTTextConfig, HunyuanViTVisionConfig
+
+    >>> # Initializing a HunyuanViTText and HunyuanViTVision configuration
+    >>> config_text = HunyuanViTTextConfig()
+    >>> config_vision = HunyuanViTVisionConfig()
+
+    >>> config = HunyuanViTConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "HunyuanViT"
+    sub_configs = {"vision_config": HunyuanViTVisionConfig}
+
+    def __init__(self, text_config=None, vision_config=None, **kwargs):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `HunyuanViTVisionConfig` with default values.")
+
+        self.vision_config = HunyuanViTVisionConfig(**vision_config)
+
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, vision_config: HunyuanViTVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`HunyuanViTConfig`] (or a derived class) from HunyuanViT text model configuration and HunyuanViT vision
+        model configuration.
+
+        Returns:
+            [`HunyuanViTConfig`]: An instance of a configuration object
+        """
+
+        return cls(vision_config=vision_config.to_dict(), **kwargs)
+
+
+__all__ = ["HunyuanViTConfig", "HunyuanViTVisionConfig"]

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ca9916451580f64b600eada187e5f4d5203cf428bacc0f4be780fe9b9527155
+size 4638523464

modeling_HunyuanViT.py

@@ -0,0 +1,838 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/HunyuanViT/modular_HunyuanViT.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_HunyuanViT.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team.
+#
+# 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 math
+import warnings
+from dataclasses import dataclass
+from typing import Any, Callable, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.utils import ModelOutput, auto_docstring, can_return_tuple, logging
+from configuration_hunyuan_vit import HunyuanViTConfig, HunyuanViTVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class HunyuanViTVisionOutput(ModelOutput):
+    """
+    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
+
+    Args:
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+@dataclass
+class HunyuanViTOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`HunyuanViTTextModel`].
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`HunyuanViTVisionModel`].
+        text_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`HunyuanViTTextModel`].
+        vision_model_output (`BaseModelOutputWithPooling`):
+            The output of the [`HunyuanViTVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: Optional[torch.FloatTensor] = None
+    logits_per_text: Optional[torch.FloatTensor] = None
+    image_embeds: Optional[torch.FloatTensor] = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class HunyuanViTVisionEmbeddings(nn.Module):
+    def __init__(self, config: HunyuanViTVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Linear(
+            in_features=config.num_channels * self.patch_size * self.patch_size,
+            out_features=self.embed_dim,
+        )
+
+        self.num_patches = config.num_patches
+        self.position_embedding_size = int(self.num_patches**0.5)
+        self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim)
+
+    @staticmethod
+    def resize_positional_embeddings(
+        positional_embeddings: torch.Tensor,
+        spatial_shapes: torch.LongTensor,
+        max_length: int,
+    ) -> torch.Tensor:
+        """
+        Resize positional embeddings to image-specific size and pad to a fixed size.
+
+        Args:
+            positional_embeddings (`torch.Tensor`):
+                Position embeddings of shape (height, width, embed_dim)
+            spatial_shapes (`torch.LongTensor`):
+                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
+            max_length (`int`):
+                Maximum length of the positional embeddings to pad resized positional embeddings to
+
+        Returns:
+            `torch.Tensor`: Embeddings of shape (batch_size, max_length, embed_dim)
+        """
+        batch_size = spatial_shapes.shape[0]
+        embed_dim = positional_embeddings.shape[-1]
+        source_dtype = positional_embeddings.dtype
+
+        resulted_positional_embeddings = torch.empty(
+            (batch_size, max_length, embed_dim),
+            device=positional_embeddings.device,
+            dtype=source_dtype,
+        )
+
+        # (height, width, embed_dim) -> (1, embed_dim, height, width) for interpolation
+        positional_embeddings = positional_embeddings.permute(2, 0, 1).unsqueeze(0)
+
+        # Upcast to float32 on CPU because antialias is not supported for bfloat16/float16 on CPU
+        if positional_embeddings.device.type == "cpu":
+            positional_embeddings = positional_embeddings.to(torch.float32)
+
+        for i in range(batch_size):
+            # (1, dim, height, width) -> (1, dim, target_height, target_width)
+            height, width = spatial_shapes[i]
+            resized_embeddings = F.interpolate(
+                positional_embeddings,
+                size=(height, width),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            )
+
+            # (1, dim, target_height, target_width) -> (target_height * target_width, dim)
+            resized_embeddings = resized_embeddings.reshape(embed_dim, height * width).transpose(0, 1)
+
+            # Cast to original dtype
+            resized_embeddings = resized_embeddings.to(source_dtype)
+
+            resulted_positional_embeddings[i, : height * width] = resized_embeddings
+            resulted_positional_embeddings[i, height * width :] = resized_embeddings[0]
+
+        return resulted_positional_embeddings
+
+    def forward(self, pixel_values: torch.FloatTensor, spatial_shapes: torch.LongTensor) -> torch.Tensor:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor`):
+                Pixel values of shape (batch_size, max_num_patches, num_channels * patch_size * patch_size)
+            spatial_shapes (`List[Tuple[int, int]]`):
+                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
+        """
+
+        # Apply patch embeddings to already patchified pixel values
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
+
+        # Get positional resized and padded positional embeddings
+        positional_embeddings = self.position_embedding.weight.reshape(
+            self.position_embedding_size, self.position_embedding_size, -1
+        )
+        resized_positional_embeddings = self.resize_positional_embeddings(
+            positional_embeddings, spatial_shapes, max_length=pixel_values.shape[1]
+        )
+
+        # Add positional embeddings to patch embeddings
+        embeddings = patch_embeds + resized_positional_embeddings
+        return embeddings
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    attn_weights = torch.matmul(query, key.transpose(-1, -2)) * scaling
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+
+    attn_output = torch.matmul(attn_weights, value)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class HunyuanViTAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, seq_length, embed_dim = hidden_states.shape
+
+        queries = self.q_proj(hidden_states)
+        keys = self.k_proj(hidden_states)
+        values = self.v_proj(hidden_states)
+
+        queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+        keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+        values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and output_attentions:
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            queries,
+            keys,
+            values,
+            attention_mask,
+            is_causal=self.is_causal,
+            scaling=self.scale,
+            dropout=0.0 if not self.training else self.dropout,
+        )
+
+        attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights
+
+
+class HunyuanViTMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class HunyuanViTEncoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Union[HunyuanViTVisionConfig]):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.self_attn = HunyuanViTAttention(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = HunyuanViTMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class HunyuanViTEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`HunyuanViTEncoderLayer`].
+
+    Args:
+        config: HunyuanViTConfig
+    """
+
+    def __init__(self, config: HunyuanViTConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([HunyuanViTEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    # Ignore copy
+    @can_return_tuple
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> BaseModelOutput:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+
+            layer_outputs = encoder_layer(
+                hidden_states,
+                attention_mask,
+                output_attentions=output_attentions,
+            )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=encoder_states,
+            attentions=all_attentions,
+        )
+
+
+class HunyuanViTVisionTransformer(nn.Module):
+    def __init__(self, config: HunyuanViTVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = HunyuanViTVisionEmbeddings(config)
+        self.encoder = HunyuanViTEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        # self.use_head = True if not hasattr(config, "vision_use_head") else config.vision_use_head
+        # if self.use_head:
+        #     self.head = HunyuanViTMultiheadAttentionPoolingHead(config)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        attention_mask: torch.Tensor,
+        spatial_shapes: torch.LongTensor,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> BaseModelOutput:
+        r"""
+        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
+            Tensor containing the spatial dimensions (height, width) of the input images.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        hidden_states = self.embeddings(pixel_values, spatial_shapes)
+
+        if attention_mask is not None and not self._use_flash_attention_2:
+            # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
+            encoder_attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+        else:
+            encoder_attention_mask = attention_mask
+
+        encoder_outputs: BaseModelOutput = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        last_hidden_state = encoder_outputs.last_hidden_state
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        # pooler_output = self.head(last_hidden_state, attention_mask) if self.use_head else None
+
+        return BaseModelOutput(
+            last_hidden_state=last_hidden_state,
+            # pooler_output=pooler_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsequently scaled and shifted by the mean and std args.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+@auto_docstring
+class HunyuanViTPreTrainedModel(PreTrainedModel):
+    config_class = HunyuanViTConfig
+    base_model_prefix = "HunyuanViT"
+    supports_gradient_checkpointing = True
+
+    _no_split_modules = [
+        "HunyuanViTTextEmbeddings",
+        "HunyuanViTEncoderLayer",
+        "HunyuanViTVisionEmbeddings",
+        "HunyuanViTEncoderLayer",
+        # "HunyuanViTMultiheadAttentionPoolingHead",
+    ]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, HunyuanViTVisionEmbeddings):
+            width = (
+                self.config.vision_config.hidden_size
+                if isinstance(self.config, HunyuanViTConfig)
+                else self.config.hidden_size
+            )
+            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
+        elif isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, HunyuanViTAttention):
+            nn.init.xavier_uniform_(module.q_proj.weight)
+            nn.init.xavier_uniform_(module.k_proj.weight)
+            nn.init.xavier_uniform_(module.v_proj.weight)
+            nn.init.xavier_uniform_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, HunyuanViTMLP):
+            nn.init.xavier_uniform_(module.fc1.weight)
+            nn.init.xavier_uniform_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        # elif isinstance(module, HunyuanViTMultiheadAttentionPoolingHead):
+        #     nn.init.xavier_uniform_(module.probe.data)
+        #     nn.init.xavier_uniform_(module.attention.in_proj_weight.data)
+        #     nn.init.zeros_(module.attention.in_proj_bias.data)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+# class HunyuanViTMultiheadAttentionPoolingHead(nn.Module):
+#     """Multihead Attention Pooling."""
+
+#     def __init__(self, config: HunyuanViTVisionConfig):
+#         super().__init__()
+
+#         self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+#         self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
+#         self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+#         self.mlp = HunyuanViTMLP(config)
+#         self.num_heads = config.num_attention_heads
+
+#     def forward(self, hidden_state: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+#         batch_size = hidden_state.shape[0]
+#         probe = self.probe.repeat(batch_size, 1, 1)
+
+#         if attention_mask is not None:
+#             target_len, source_len = probe.shape[1], hidden_state.shape[1]
+#             attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_state.dtype, target_len)
+#             attention_mask = attention_mask.repeat(1, self.num_heads, target_len, 1)
+#             attention_mask = attention_mask.reshape(-1, target_len, source_len)
+
+#         hidden_state = self.attention(probe, hidden_state, hidden_state, attn_mask=attention_mask)[0]
+
+#         residual = hidden_state
+#         hidden_state = self.layernorm(hidden_state)
+#         hidden_state = residual + self.mlp(hidden_state)
+
+#         return hidden_state[:, 0]
+
+
+@auto_docstring(
+    custom_intro="""
+    The vision model from HunyuanViT without any head or projection on top.
+    """
+)
+class HunyuanViTVisionModel(HunyuanViTPreTrainedModel):
+    config_class = HunyuanViTVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: HunyuanViTVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = HunyuanViTVisionTransformer(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        pixel_attention_mask: torch.Tensor,
+        spatial_shapes: torch.LongTensor,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> BaseModelOutputWithPooling:
+        r"""
+        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
+            Mask to avoid performing attention on padding pixel indices.
+        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
+            Tensor containing the spatial dimensions (height, width) of the input images.
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, HunyuanViTVisionModel
+
+        >>> model = HunyuanViTVisionModel.from_pretrained("google/HunyuanViT-base-patch16-224")
+        >>> processor = AutoProcessor.from_pretrained("google/HunyuanViT-base-patch16-224")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled features
+        ```"""
+        return self.vision_model(
+            pixel_values=pixel_values,
+            attention_mask=pixel_attention_mask,
+            spatial_shapes=spatial_shapes,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+
+@auto_docstring
+class HunyuanViTModel(HunyuanViTPreTrainedModel):
+    config_class = HunyuanViTConfig
+
+    def __init__(self, config: HunyuanViTConfig):
+        super().__init__(config)
+
+        if not isinstance(config.vision_config, HunyuanViTVisionConfig):
+            raise TypeError(
+                "config.vision_config is expected to be of type HunyuanViTVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        vision_config = config.vision_config
+
+        # First, initialize the text and vision models with proper attention implementation
+        vision_model = HunyuanViTVisionModel._from_config(vision_config)
+
+        # Second, get the text and vision submodules (for backward compatibility)
+        self.vision_model = vision_model.vision_model
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        pixel_attention_mask: Optional[torch.Tensor] = None,
+        spatial_shapes: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
+            Mask to avoid performing attention on padding pixel indices.
+        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
+            Tensor containing the spatial dimensions (height, width) of the input images.
+
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`HunyuanViTVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AutoModel
+        >>> import torch
+
+        >>> model = AutoModel.from_pretrained("google/HunyuanViT-base-patch16-224")
+        >>> processor = AutoProcessor.from_pretrained("google/HunyuanViT-base-patch16-224")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> with torch.no_grad():
+        ...     image_features = model.get_image_features(**inputs)
+        ```
+        """
+        # Use HunyuanViTModel's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
+            pixel_values=pixel_values,
+            attention_mask=pixel_attention_mask,
+            spatial_shapes=spatial_shapes,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = vision_outputs.pooler_output
+
+        return pooled_output
+
+
+__all__ = [
+    "HunyuanViTModel",
+    "HunyuanViTPreTrainedModel",
+    "HunyuanViTVisionModel",
+    "HunyuanViTForImageClassification",
+]

preprocessor_config.json

@@ -0,0 +1,23 @@
+{
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "HunyuanViTImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "processor_class": "HunyuanViTProcessor",
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "base_size": 0,
+  "maxres": 1536,
+  "minres": 0
+}