first commit

README.md

@@ -1,3 +1,53 @@
 ---
 license: cc-by-nc-4.0
 ---
+
+# QLIP
+
+[\[📂 GitHub\]](https://github.com/NVlabs/QLIP)
+[\[📃 QLIP Tech Report\]](http://arxiv.org/abs/2502.yyyyy)
+[\[🔗 Project Page\]](http://nvlabs.github.io/QLIP/)
+[\[🤗 HF Model\]](https://huggingface.co/NVIDIA/QLIP-L-14-392)
+
+## Introduction
+We introduce Quantized Language-Image Pretraining (**QLIP**), a visual tokenization method that combines state-of-the-art reconstruction quality with state-of-the-art zero-shot image understanding.
+QLIP trains a binary-spherical-quantization-based autoencoder with reconstruction and language-image alignment objectives.
+We are the first to show that the two objectives do not need to be at odds.
+We balance the two loss terms dynamically during training and show that a two-stage training pipeline effectively mixes the large-batch requirements of image-language pre-training with the memory bottleneck imposed by the reconstruction objective.
+We validate the effectiveness of QLIP for multimodal understanding and text-conditioned image generation with a single model.
+Specifically, QLIP serves as a drop-in replacement for the visual encoder for LLaVA and the image tokenizer for LlamaGen with comparable or even better performance.
+Finally, we demonstrate that QLIP enables a unified mixed-modality auto-regressive model for understanding and generation.
+
+## Model Zoo
+We provide the following models:
+| model name    | #bits  |  CR<sub>&uarr;<sub>   | 0-shot<sub>&uarr;<sub> | rFID<sub>&darr;<sub> | HF Link |
+| ------------- | ------ | ----- | ------ | ---- | ------- |
+| QLIP-B-16-256 |   28   | 219.4 |  74.3  | 3.21 | [🤗 link](https://huggingface.co/NVIDIA/QLIP-B-16-256) |
+| QLIP-B-8-256  |   28   |  54.8 |  75.6  | 0.70 | [🤗 link](https://huggingface.co/NVIDIA/QLIP-B-8-256)  |
+| QLIP-L-14-392 |   28   | 168   |  79.1  | 1.46 | [🤗 link](https://huggingface.co/NVIDIA/QLIP-L-14-392) |
+
+Note:  
+- **CR**: compression ratio = 24/(#bits)*patch_size^2;
+- **0-shot**: zero-shot classification accuracy on IN-1k-val;
+- **rFID**: reconstruction FID on IN-1k-val.
+
+## Citing QLIP
+
+```bibtex
+@article{zhao2025qlip,
+  title={QLIP: Text-Aligned Visual Tokenization Unifies Auto-Regressive Multimodal Understanding and Generation},
+  author={Zhao, Yue and Xue, Fuzhao and Reed, Scott and Fan, Linxi and Zhu, Yuke and Kautz, Jan and Yu, Zhiding and Krähenbühl, Philipp and Huang, De-An},
+  journal={arXiv preprint arXiv:2502.yyyyy},
+  year={2025}
+}
+```
+
+## Acknowledgement
+The project builds upon the following open-source efforts:
+- [EVA-CLIP](https://github.com/baaivision/EVA/tree/master/EVA-CLIP/rei): We use EVA-CLIP as initialization which significantly speeds up the training convergence.
+
+- [LLaVA](https://github.com/haotian-liu/LLaVA): We use LLaVA to evaluate the multimodal understanding performance.
+
+- [LlamaGen](https://github.com/FoundationVision/LlamaGen): We build the text-to-image generation evaluation on top of LlamaGen.
+
+- [Lingua](https://github.com/facebookresearch/lingua): We build the unified multimodal model on top of Lingua.

bsq.py

@@ -0,0 +1,227 @@
+# Copyright (c) 2024, NVIDIA Corporation & Affiliates. All rights reserved. 
+# 
+# This work is made available under the Nvidia Source Code License-NC. 
+# To view a copy of this license, visit 
+# https://github.com/NVlabs/QLIP/blob/main/LICENSE
+
+# MIT License
+# Based on https://github.com/zhaoyue-zephyrus/bsq-vit/blob/main/transcoder/models/quantizer/bsq.py
+
+import torch
+import torch.nn as nn
+from einops import rearrange, reduce
+
+_EPS = 1e-8
+
+
+class DifferentiableEntropyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, zq, basis, K, eps):
+        zb = (zq + 1) / 2
+        zi = ((zb * basis).sum(-1)).to(torch.int64)
+        cnt = torch.scatter_reduce(
+            torch.zeros(2**K, device=zq.device, dtype=zq.dtype),
+            0,
+            zi.flatten(),
+            torch.ones_like(zi.flatten()).to(zq.dtype),
+            "sum",
+        )
+        prob = (cnt + eps) / (cnt + eps).sum()
+        H = torch.special.entr(prob).sum()
+        ctx.save_for_backward(zq, zi, prob)
+        ctx.K = K
+        return H
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        zq, zi, prob = ctx.saved_tensors
+        grad_array = -grad_output * (torch.log(prob) + 1) / zi.numel() / ctx.K
+        reord_grad = grad_array[zi.flatten()].reshape(zi.shape)
+        grad_input = reord_grad.unsqueeze(-1) * zq
+        return grad_input, None, None, None, None
+
+
+def codebook_entropy(zq, basis, K, eps=1e-8):
+    return DifferentiableEntropyFunction.apply(zq, basis, K, eps)
+
+
+class BinarySphericalQuantizer(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int = 18,
+        group_size: int = 9,
+        soft_entropy: bool = True,
+        beta: float = 0.0,  # commit loss
+        gamma_0: float = 1.0,  # entropy loss (E[H(q)])
+        gamma_1: float = 1.0,  # entropy loss (H[E[q]])
+        input_format: str = "bchw",
+        persample_entropy_compute: str = "group",
+        l2_norm: bool = True,
+        inv_temperature: float = 100.0,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.group_size = group_size
+        assert embed_dim % group_size == 0, "embed_dim must be divisible by group_size"
+        self.soft_entropy = soft_entropy
+        self.beta = beta
+        self.gamma_0 = gamma_0
+        self.gamma_1 = gamma_1
+        assert input_format in ["bchw", "blc"]
+        self.input_format = input_format
+        assert persample_entropy_compute in [
+            "group",
+            "analytical",
+        ], "persample_entropy_compute must be either 'group' or 'analytical'"
+        self.persample_entropy_compute = persample_entropy_compute
+        self.l2_norm = l2_norm
+        self.inv_temperature = inv_temperature
+
+        self.register_buffer("basis", 2 ** torch.arange(embed_dim - 1, -1, -1), persistent=False)
+        self.register_buffer(
+            "group_basis", 2 ** torch.arange(group_size - 1, -1, -1), persistent=False
+        )
+
+        group_codes = torch.arange(2**self.group_size)
+        group_codebook = self.indexes_to_codes(group_codes).float()[:, -group_size:]
+        self.register_buffer("group_codebook", group_codebook, persistent=False)
+
+    def quantize(self, z):
+        assert (
+            z.shape[-1] == self.embed_dim
+        ), f"Expected {self.embed_dim} dimensions, got {z.shape[-1]}"
+        zhat = torch.where(z > 0, torch.ones_like(z), -torch.ones_like(z))
+        return z + (zhat - z).detach()
+
+    def forward(self, z):
+        if self.input_format == "bchw":
+            z = rearrange(z, "b c h w -> b h w c")
+        zq = self.quantize(z)
+
+        indices = self.codes_to_indexes(zq.detach())
+        group_indices = self.codes_to_group_indexes(zq.detach())
+
+        if not self.training:
+            used_codes = torch.unique(indices, return_counts=False)
+        else:
+            used_codes = None
+
+        if self.soft_entropy:
+            persample_entropy, cb_entropy = self.soft_entropy_loss(z)
+        else:
+            persample_entropy, cb_entropy = self.hard_entropy_loss(z)
+        entropy_penalty = self.gamma_0 * persample_entropy - self.gamma_1 * cb_entropy
+
+        q_scale = 1.0 / (self.embed_dim**0.5) if self.l2_norm else 1.0
+        zq = zq * q_scale
+        commit_loss = self.beta * torch.mean(((zq.detach() - z) ** 2).sum(dim=-1))
+
+        if self.input_format == "bchw":
+            zq = rearrange(zq, "b h w c -> b c h w")
+
+        return (
+            zq,
+            commit_loss + entropy_penalty / self.inv_temperature,
+            {
+                "H": cb_entropy,
+                "used_codes": used_codes,
+                "indices": indices,
+                "group_indices": group_indices,
+            },
+        )
+
+    def soft_entropy_loss(self, z):
+        group_codebook = self.group_codebook / (self.embed_dim**0.5 if self.l2_norm else 1)
+        divided_z = rearrange(z, "... (g c) -> ... g c", c=self.group_size)
+
+        if self.persample_entropy_compute == "group":
+            distance = -2 * torch.einsum("... g c, d c -> ... g d", divided_z, group_codebook)
+            prob = (-distance * self.inv_temperature).softmax(dim=-1)
+            persample_entropy = torch.special.entr(prob + _EPS).sum((-1, -2)).mean()
+        else:
+            p = torch.sigmoid(
+                -4 * z / (self.embed_dim**0.5 if self.l2_norm else 1) * self.inv_temperature
+            )
+            prob = torch.stack([p, 1 - p], dim=-1)
+            persample_entropy = torch.special.entr(prob + _EPS).sum((-1, -2)).mean()
+
+        # macro average of the probability of each subgroup
+        avg_prob = reduce(prob, "... g d -> g d", "mean")
+        cb_entropy = torch.special.entr(avg_prob + _EPS).sum()
+
+        return persample_entropy, cb_entropy
+
+    def hard_entropy_loss(self, z):
+        zb = ((z + 1) / 2).reshape(z.shape[0], -1, z.shape[-1]).to(torch.float32)
+        prob_per_dim = zb.sum(1) / zb.shape[1]
+        prob = torch.stack([prob_per_dim, 1 - prob_per_dim], dim=-1)
+        persample_entropy = torch.special.entr(prob + _EPS).sum((-1, -2)).mean()
+        cb_entropy = codebook_entropy(z, self.basis, self.embed_dim)
+
+        return persample_entropy, cb_entropy
+
+    def codes_to_indexes(self, zhat):
+        """Converts a `code` to an index in the codebook.
+        Args:
+            zhat: A tensor of shape (B, ..., C) containing the codes. must be in {-1, 1}
+        """
+        assert (
+            zhat.shape[-1] == self.embed_dim
+        ), f"Expected {self.embed_dim} dimensions, got {zhat.shape[-1]}"
+        return ((zhat.int() + 1) / 2 * self.basis).sum(axis=-1).to(torch.int64)
+
+    def codes_to_group_indexes(self, zhat):
+        """Converts a `code` to a list of indexes (in groups) in the codebook.
+        Args:
+            zhat: A tensor of shape (B, ..., C) containing the codes. must be in {-1, 1}
+        """
+        zhat_in_group = rearrange(zhat, "b ... (g c) -> b ... g c", c=self.group_size)
+        return ((zhat_in_group.int() + 1) / 2 * self.group_basis).sum(axis=-1).to(torch.int64)
+
+    def indexes_to_codes(self, indices):
+        """Inverse of `codes_to_indexes`."""
+        indices = indices.unsqueeze(-1)
+        codes_non_centered = torch.remainder(torch.floor_divide(indices, self.basis), 2)
+        return codes_non_centered * 2 - 1
+
+    def group_indexes_to_codes(self, group_indices):
+        """Inverse of `codes_to_group_indexes`."""
+        group_indices = group_indices.unsqueeze(-1)
+        codes_non_centered = torch.remainder(torch.floor_divide(group_indices, self.group_basis), 2)
+        codes_non_centered = rearrange(codes_non_centered, "b ... g c -> b ... (g c)")
+        return codes_non_centered * 2 - 1
+
+    def get_group_codebook_entry(self, group_indices, one_hot=False):
+        """
+        Args:
+            group_indices: A tensor of shape (B, L, G, C) containing the group indices.
+        """
+        if one_hot:
+            z_q = group_indices @ self.group_codebook
+        else:
+            z_q = self.group_indexes_to_codes(group_indices)
+        q_scale = 1.0 / (self.embed_dim**0.5) if self.l2_norm else 1.0
+        z_q = z_q * q_scale
+        if self.input_format == "bchw":
+            h, w = int(z_q.shape[1] ** 0.5)
+            assert h * w == z_q.shape[1], "Invalid sequence length"
+            z_q = rearrange(z_q, "b (h w) c -> b c h w", h=h)
+        return z_q
+
+    def get_codebook_entry(self, indices, one_hot=False):
+        """
+        Args:
+            group_indices: A tensor of shape (B, L, C) containing the indices.
+        """
+        if one_hot:
+            assert self.embed_dim == self.group_size, "one_hot is only supported for group_size == embed_dim"
+            z_q = indices @ self.group_codebook
+        else:
+            z_q = self.indexes_to_codes(indices)
+        q_scale = 1.0 / (self.embed_dim**0.5) if self.l2_norm else 1.0
+        z_q = z_q * q_scale
+        if self.input_format == "bchw":
+            h, w = int(z_q.shape[1] ** 0.5)
+            assert h * w == z_q.shape[1], "Invalid sequence length"
+            z_q = rearrange(z_q, "b (h w) c -> b c h w", h=h)
+        return z_q

config.json

@@ -0,0 +1,79 @@
+{
+    "_name_or_path": "EVA-BSQCLIP",
+    "architectures": [
+      "QLIPModel"
+    ],
+    "auto_map": {
+      "AutoConfig": "configuration_evaclip.QLIPConfig",
+      "AutoModel": "modeling_evaclip.QLIPModel"
+    },
+    "decoder_config": {
+      "dropout": 0.0,
+      "hidden_size": 1024,
+      "hidden_size_post_q": 1024,
+      "image_size": 392,
+      "intermediate_size": 2730,
+      "k_bias": false,
+      "layer_norm_eps": 1e-06,
+      "model_type": "clip_decoder_model",
+      "num_attention_heads": 16,
+      "num_hidden_layers": 24,
+      "patch_size": 14,
+      "rope": true,
+      "rope_shift": 0,
+      "subln": true,
+      "swiglu": true,
+      "use_bfloat16": true,
+      "use_rms_norm": true
+    },
+    "initializer_factor": 1.0,
+    "logit_scale_init_value": 2.6592,
+    "model_type": "clip",
+    "projection_dim": 768,
+    "text_config": {
+      "bos_token_id": 0,
+      "dropout": 0.0,
+      "eos_token_id": 2,
+      "hidden_size": 768,
+      "intermediate_size": 3072,
+      "num_attention_heads": 12,
+      "model_type": "clip_text_model",
+      "use_bfloat16": true,
+      "use_rms_norm": false
+    },
+    "text_projection_bias": false,
+    "torch_dtype": "float32",
+    "transformers_version": "4.37.2",
+    "vision_config": {
+      "dropout": 0.0,
+      "hidden_size": 1024,
+      "hidden_size_post_q": 1024,
+      "image_size": 392,
+      "intermediate_size": 2730,
+      "k_bias": false,
+      "layer_norm_eps": 1e-06,
+      "model_type": "clip_vision_model",
+      "num_attention_heads": 16,
+      "num_hidden_layers": 24,
+      "patch_size": 14,
+      "projection_dim": 1024,
+      "quantizer": "bsq",
+      "quantizer_cfg": {
+        "embed_dim": 28,
+        "group_size": 1,
+        "input_format": "blc",
+        "inv_temperature": 1.0,
+        "l2_norm": true
+      },
+      "quantizer_embed_type": "mlp",
+      "quantizer_l2_norm": true,
+      "rope": true,
+      "rope_shift": 1,
+      "subln": true,
+      "swiglu": true,
+      "use_bfloat16": true,
+      "use_rms_norm": true
+    },
+    "vision_projection_bias": true
+  }
+  

configuration_qlip.py

@@ -0,0 +1,566 @@
+# Copyright (c) 2024, NVIDIA Corporation & Affiliates. All rights reserved. 
+# 
+# This work is made available under the Nvidia Source Code License-NC. 
+# To view a copy of this license, visit 
+# https://github.com/NVlabs/QLIP/blob/main/LICENSE
+
+# Copyright 2021 The HuggingFace Inc. 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.
+""" CLIP model configuration"""
+
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import ProcessorMixin
+    from transformers.utils import TensorType
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class QLIPTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPTextModel`]. It is used to instantiate a CLIP
+    text 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 text encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 49408):
+            Vocabulary size of the CLIP text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`CLIPModel`].
+        hidden_size (`int`, *optional*, defaults to 512):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        max_position_embeddings (`int`, *optional*, defaults to 77):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            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-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+        pad_token_id (`int`, *optional*, defaults to 1):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 49406):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 49407):
+            End of stream token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPTextConfig, CLIPTextModel
+
+    >>> # Initializing a CLIPTextConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPTextConfig()
+
+    >>> # Initializing a CLIPTextModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clip_text_model"
+
+    def __init__(
+        self,
+        vocab_size=49408,
+        hidden_size=512,
+        intermediate_size=2048,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=8,
+        max_position_embeddings=77,
+        hidden_act="gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        # This differs from `CLIPTokenizer`'s default and from openai/clip
+        # See https://github.com/huggingface/transformers/pull/24773#issuecomment-1632287538
+        q_bias=True,
+        k_bias=True,
+        v_bias=True,
+        subln=False,
+        swiglu=False,
+        rope=False,
+        post_layernorm=False,
+        pad_token_id=1,
+        bos_token_id=49406,
+        eos_token_id=49407,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.q_bias=q_bias
+        self.k_bias=k_bias
+        self.v_bias=v_bias
+        self.subln = subln
+        self.swiglu = swiglu
+        self.rope = rope
+        self.post_layernorm = post_layernorm
+        self.attention_dropout = attention_dropout
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class QLIPVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
+    CLIP 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 CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) 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.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        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):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            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-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPVisionConfig, CLIPVisionModel
+
+    >>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPVisionConfig()
+
+    >>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        q_bias=True,
+        k_bias=True,
+        v_bias=True,
+        subln=False,
+        swiglu=False,
+        rope=False,
+        post_layernorm=False,
+        # quantizer specs
+        quantizer="none",
+        quantizer_l2_norm=False,
+        quantizer_embed_type="identity",
+        hidden_size_post_q=None,
+        quantizer_cfg=dict(),
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        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.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.q_bias=q_bias
+        self.k_bias=k_bias
+        self.v_bias=v_bias
+        self.subln = subln
+        self.swiglu = swiglu
+        self.rope = rope
+        self.post_layernorm = post_layernorm
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+        self.quantizer = quantizer
+        self.quantizer_l2_norm = quantizer_l2_norm
+        self.quantizer_embed_type = quantizer_embed_type
+        self.hidden_size_post_q = self.hidden_size if hidden_size_post_q is None else hidden_size_post_q
+        self.quantizer_cfg = quantizer_cfg
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class QLIPDecoderConfig(PretrainedConfig):
+    model_type = "clip_decoder_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        q_bias=True,
+        k_bias=True,
+        v_bias=True,
+        subln=False,
+        swiglu=False,
+        rope=False,
+        post_layernorm=False,
+        # quantizer specs
+        quantizer="none",
+        quantizer_l2_norm=False,
+        quantizer_embed_type="identity",
+        hidden_size_post_q=None,
+        quantizer_cfg=dict(),
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        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.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.q_bias=q_bias
+        self.k_bias=k_bias
+        self.v_bias=v_bias
+        self.subln = subln
+        self.swiglu = swiglu
+        self.rope = rope
+        self.post_layernorm = post_layernorm
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+        self.quantizer = quantizer
+        self.quantizer_l2_norm = quantizer_l2_norm
+        self.quantizer_embed_type = quantizer_embed_type
+        self.hidden_size_post_q = self.hidden_size if hidden_size_post_q is None else hidden_size_post_q
+        self.quantizer_cfg = quantizer_cfg
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class QLIPConfig(PretrainedConfig):
+    r"""
+    [`CLIPConfig`] is the configuration class to store the configuration of a [`CLIPModel`]. It is used to instantiate
+    a CLIP 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 CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) 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 [`CLIPTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original CLIP implementation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPConfig, CLIPModel
+
+    >>> # Initializing a CLIPConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPConfig()
+
+    >>> # Initializing a CLIPModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a CLIPConfig from a CLIPTextConfig and a CLIPVisionConfig
+    >>> from transformers import CLIPTextConfig, CLIPVisionConfig
+
+    >>> # Initializing a CLIPText and CLIPVision configuration
+    >>> config_text = CLIPTextConfig()
+    >>> config_vision = CLIPVisionConfig()
+
+    >>> config = CLIPConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "clip"
+
+    def __init__(
+        self, text_config=None, vision_config=None, decoder_config=None, projection_dim=512, logit_scale_init_value=2.6592, **kwargs
+    ):
+        # If `_config_dict` exist, we use them for the backward compatibility.
+        # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
+        # of confusion!).
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+        decoder_config_dict = kwargs.pop("decoder_config_dict", None)
+
+        super().__init__(**kwargs)
+
+        # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
+        # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
+        # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
+        if text_config_dict is not None:
+            if text_config is None:
+                text_config = {}
+
+            # This is the complete result when using `text_config_dict`.
+            _text_config_dict = QLIPTextConfig(**text_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
+            for key, value in _text_config_dict.items():
+                if key in text_config and value != text_config[key] and key not in ["transformers_version"]:
+                    # If specified in `text_config_dict`
+                    if key in text_config_dict:
+                        message = (
+                            f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
+                            f'The value `text_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The "
+                            f'value `text_config["{key}"]` will be overriden.'
+                        )
+                    logger.info(message)
+
+            # Update all values in `text_config` with the ones in `_text_config_dict`.
+            text_config.update(_text_config_dict)
+
+        if vision_config_dict is not None:
+            if vision_config is None:
+                vision_config = {}
+
+            # This is the complete result when using `vision_config_dict`.
+            _vision_config_dict = QLIPVisionConfig(**vision_config_dict).to_dict()
+            # convert keys to string instead of integer
+            if "id2label" in _vision_config_dict:
+                _vision_config_dict["id2label"] = {
+                    str(key): value for key, value in _vision_config_dict["id2label"].items()
+                }
+
+            # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
+            for key, value in _vision_config_dict.items():
+                if key in vision_config and value != vision_config[key] and key not in ["transformers_version"]:
+                    # If specified in `vision_config_dict`
+                    if key in vision_config_dict:
+                        message = (
+                            f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
+                            f'values. The value `vision_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`vision_config_dict` is provided which will be used to initialize `CLIPVisionConfig`. "
+                            f'The value `vision_config["{key}"]` will be overriden.'
+                        )
+                    logger.info(message)
+
+            # Update all values in `vision_config` with the ones in `_vision_config_dict`.
+            vision_config.update(_vision_config_dict)
+
+        if decoder_config_dict is not None:
+            if decoder_config is None:
+                decoder_config = {}
+
+            # This is the complete result when using `decoder_config_dict`.
+            _decoder_config_dict = QLIPDecoderConfig(**decoder_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_decoder_config_dict` and `decoder_config` but being different.
+            for key, value in _decoder_config_dict.items():
+                if key in decoder_config and value != decoder_config[key] and key not in ["transformers_version"]:
+                    # If specified in `decoder_config_dict`
+                    if key in decoder_config_dict:
+                        message = (
+                            f"`{key}` is found in both `decoder_config_dict` and `decoder_config` but with different values. "
+                            f'The value `decoder_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`decoder_config_dict` is provided which will be used to initialize `QLIPDecoderConfig`. The "
+                            f'value `decoder_config["{key}"]` will be overriden.'
+                        )
+                    logger.info(message)
+
+            # Update all values in `decoder_config` with the ones in `_decoder_config_dict`.
+            decoder_config.update(_decoder_config_dict)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `CLIPTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `CLIPVisionConfig` with default values.")
+
+        if decoder_config is None:
+            decoder_config = {}
+            logger.info("`decoder_config` is `None`. initializing the `CLIPDecoderConfig` with default values.")
+
+        self.text_config = QLIPTextConfig(**text_config)
+        self.vision_config = QLIPVisionConfig(**vision_config)
+        self.decoder_config = QLIPDecoderConfig(**decoder_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: QLIPTextConfig, vision_config: QLIPVisionConfig, decoder_config: QLIPDecoderConfig, **kwargs):
+        r"""
+        Instantiate a [`CLIPConfig`] (or a derived class) from clip text model configuration and clip vision model
+        configuration.
+
+        Returns:
+            [`CLIPConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), decoder_config=decoder_config.to_dict(), **kwargs)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edb2e10b0fdab7aab3a4019a019fec980b963685cf78d0c1e21179b58346cf9e
+size 2953399476

modeling_qlip.py

@@ -0,0 +1,1482 @@
+# Copyright (c) 2024, NVIDIA Corporation & Affiliates. All rights reserved. 
+# 
+# This work is made available under the Nvidia Source Code License-NC. 
+# To view a copy of this license, visit 
+# https://github.com/NVlabs/QLIP/blob/main/LICENSE
+
+# coding=utf-8
+# Copyright 2021 The OpenAI Team 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.
+""" PyTorch CLIP model."""
+
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+from einops import rearrange
+import torch
+import torch.utils.checkpoint
+from torch import nn
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+
+from configuration_qlip import QLIPConfig, QLIPTextConfig, QLIPVisionConfig, QLIPDecoderConfig
+from bsq import BinarySphericalQuantizer
+from rope import VisionRotaryEmbeddingFast
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"
+
+CLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "openai/clip-vit-base-patch32",
+    # See all CLIP models at https://huggingface.co/models?filter=clip
+]
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/2021-03-07-clip.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+class QLIPVisionModelOutput(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: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class QLIPTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The text 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.
+    """
+
+    text_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class QLIPOutput(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 [`CLIPTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPVisionModel`].
+        text_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`CLIPTextModel`].
+        vision_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`CLIPVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPooling = None
+    vision_model_output: BaseModelOutputWithPooling = None
+    reconstructions: torch.FloatTensor = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class QLIPVisionEmbeddings(nn.Module):
+    def __init__(self, config: QLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=True,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class QLIPTextEmbeddings(nn.Module):
+    def __init__(self, config: QLIPTextConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if inputs_embeds is None:
+            inputs_embeds = self.token_embedding(input_ids)
+
+        position_embeddings = self.position_embedding(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+
+        return embeddings
+
+
+class QLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config, rope=None, rope_shift=1):
+        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.subln = config.subln
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.k_bias)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.v_bias)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.q_bias)
+        self.inner_attn_ln = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) if config.subln else nn.Identity()
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)
+
+        self.rope = rope
+        self.rope_shift = rope_shift
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scale
+        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        if self.rope:
+            q_t = query_states[:, self.rope_shift:, :]
+            ro_q_t = self.rope(q_t)
+            query_states = torch.cat([query_states[:, :self.rope_shift, :], ro_q_t], dim=-2).type_as(value_states)
+
+            k_t = key_states[:, self.rope_shift:, :]
+            ro_k_t = self.rope(k_t)
+            key_states = torch.cat([key_states[:, :self.rope_shift, :], ro_k_t], dim=-2).type_as(value_states)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        # apply the causal_attention_mask first
+        if causal_attention_mask is not None:
+            if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {causal_attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        attn_output = self.inner_attn_ln(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped
+
+
+class QLIPSwiGLU(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.w1 = nn.Linear(self.hidden_size, self.intermediate_size)
+        self.w2 = nn.Linear(self.hidden_size, self.intermediate_size)
+        self.w3 = nn.Linear(self.intermediate_size, self.hidden_size)
+        self.act_fn = nn.SiLU()
+        self.ffn_ln = nn.LayerNorm(self.intermediate_size, eps=config.layer_norm_eps) if config.subln else nn.Identity()
+
+    def forward(self, x):
+        x1 = self.w1(x)
+        x2 = self.w2(x)
+        hidden = self.act_fn(x1) * x2
+        x = self.ffn_ln(hidden)
+        x = self.w3(x)
+        return x
+
+
+class QLIPMLP(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)
+        self.ffn_ln = nn.LayerNorm(config.intermediate_size, eps=config.layer_norm_eps) if config.subln else nn.Identity()
+
+    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.ffn_ln(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class QLIPEncoderLayer(nn.Module):
+    def __init__(self, config: QLIPConfig, rope=None, rope_shift=1):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = QLIPAttention(config, rope=rope, rope_shift=rope_shift)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = QLIPSwiGLU(config) if config.swiglu else QLIPMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        causal_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 size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                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,
+            causal_attention_mask=causal_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 QLIPPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = QLIPConfig
+    base_model_prefix = "clip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, QLIPTextEmbeddings):
+            module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+        elif isinstance(module, QLIPVisionEmbeddings):
+            factor = self.config.initializer_factor
+            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
+            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
+            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+        elif isinstance(module, QLIPAttention):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            out_proj_std = (module.embed_dim**-0.5) * factor
+            nn.init.normal_(module.q_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.k_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.v_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
+        elif isinstance(module, QLIPMLP):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
+            nn.init.normal_(module.fc1.weight, std=fc_std)
+            nn.init.normal_(module.fc2.weight, std=in_proj_std)
+        elif isinstance(module, QLIPModel):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, QLIPVisionModelWithProjection):
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+        elif isinstance(module, QLIPTextModelWithProjection):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
+            )
+
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+CLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`CLIPConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        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)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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.
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        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.
+"""
+
+CLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        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)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        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.
+"""
+
+
+class QLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`CLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(self, config: QLIPConfig, rope=None, rope_shift=1):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([
+            QLIPEncoderLayer(config, rope=rope, rope_shift=rope_shift)
+            for _ in range(config.num_hidden_layers)
+        ])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, 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)
+            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Causal mask for the text model. 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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    causal_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,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class QLIPTextTransformer(nn.Module):
+    def __init__(self, config: QLIPTextConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.embeddings = QLIPTextEmbeddings(config)
+        self.encoder = QLIPEncoder(config)
+        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=QLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is None:
+            raise ValueError("You have to specify input_ids")
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = _create_4d_causal_attention_mask(
+            input_shape, hidden_states.dtype, device=hidden_states.device
+        )
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+            ]
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                # We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
+                (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                .int()
+                .argmax(dim=-1),
+            ]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """The text model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class QLIPTextModel(QLIPPreTrainedModel):
+    config_class = QLIPTextConfig
+
+    _no_split_modules = ["QLIPTextEmbeddings", "QLIPEncoderLayer"]
+
+    def __init__(self, config: QLIPTextConfig):
+        super().__init__(config)
+        self.text_model = QLIPTextTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=QLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPTextModel
+
+        >>> model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class QLIPVisionTransformer(nn.Module):
+    def __init__(self, config: QLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = QLIPVisionEmbeddings(config)
+        if config.rope:
+            half_head_dim = config.hidden_size // config.num_attention_heads // 2
+            hw_seq_len = config.image_size // config.patch_size
+            self.rope = VisionRotaryEmbeddingFast(
+                dim=half_head_dim,
+                pt_seq_len=16,
+                ft_seq_len=hw_seq_len,
+            )
+        else:
+            self.rope = None
+        self.encoder = QLIPEncoder(config, rope=self.rope, rope_shift=1)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        if config.quantizer == "bsq":
+            self.quantizer = BinarySphericalQuantizer(**config.quantizer_cfg)
+        self.quantizer_l2_norm = config.quantizer_l2_norm
+        if config.quantizer_embed_type == "mlp":
+            self.quant_embed = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("c_fc", nn.Linear(config.hidden_size, config.hidden_size)),
+                        ("gelu", nn.GELU()),
+                        ("c_proj", nn.Linear(config.hidden_size, config.quantizer_cfg["embed_dim"])),
+                    ]
+                )
+            )
+            self.quant_embed_post = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("c_fc", nn.Linear(config.quantizer_cfg["embed_dim"], config.hidden_size_post_q)),
+                        ("gelu", nn.GELU()),
+                        ("c_proj", nn.Linear(config.hidden_size_post_q, config.hidden_size_post_q)),
+                    ]
+                )
+            )
+        else:
+            self.quant_embed = nn.Identity()
+            self.quant_embed_post = nn.Identity()
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=QLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = last_hidden_state[:, 0, :]
+        z = last_hidden_state[:, 1:, :]
+        h = self.quant_embed(z)
+        if self.quantizer_l2_norm:
+            h = F.normalize(h, dim=-1)
+        if self.quantizer is not None:
+            quant, _, _ = self.quantizer(h)
+        else:
+            quant = h
+        zhat = self.quant_embed_post(quant)
+        last_hidden_state = zhat
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class QLIPVisionTransformerDecoder(nn.Module):
+    def __init__(self, config: QLIPDecoderConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        num_patches = (config.image_size // config.patch_size) ** 2
+        self.patch_shape = (config.image_size // config.patch_size, config.image_size // config.patch_size)
+        self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+        if config.rope:
+            half_head_dim = config.hidden_size // config.num_attention_heads // 2
+            hw_seq_len = config.image_size // config.patch_size
+            self.rope = VisionRotaryEmbeddingFast(
+                dim=half_head_dim,
+                pt_seq_len=16,
+                ft_seq_len=hw_seq_len,
+            )
+        else:
+            self.rope = None
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.encoder = QLIPEncoder(config, rope=self.rope, rope_shift=0)
+        self.ffn = nn.Sequential(
+            nn.Linear(config.hidden_size, config.intermediate_size),
+            nn.Tanh(),
+        )
+        self.conv_out = nn.Linear(
+            in_features=config.intermediate_size,
+            out_features=3 * config.patch_size * config.patch_size,
+        )
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=QLIPVisionConfig)
+    def forward(
+        self,
+        latents: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if latents is None:
+            raise ValueError("You have to specify latents")
+
+        hidden_states = self.position_embeddings + latents
+
+        decoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = decoder_outputs[0]
+        recon = self.conv_out(self.ffn(self.norm(last_hidden_state)))
+        recon_reshaped = rearrange(
+            recon, "b (hh ww) (c sh sw) -> b c (hh sh) (ww sw)",
+            hh=self.patch_shape[0], ww=self.patch_shape[1],
+            sh=self.config.patch_size, sw=self.config.patch_size,
+        )
+        return recon_reshaped
+
+
+@add_start_docstrings(
+    """The vision model from CLIP without any head or projection on top.""",
+    CLIP_START_DOCSTRING,
+)
+class QLIPVisionModel(QLIPPreTrainedModel):
+    config_class = QLIPVisionConfig
+    main_input_name = "pixel_values"
+    _no_split_modules = ["QLIPEncoderLayer"]
+
+    def __init__(self, config: QLIPVisionConfig):
+        super().__init__(config)
+        self.vision_model = QLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=QLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPVisionModel
+
+        >>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> 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 CLS states
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+@add_start_docstrings(CLIP_START_DOCSTRING)
+class QLIPModel(QLIPPreTrainedModel):
+    config_class = QLIPConfig
+
+    def __init__(self, config: QLIPConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, QLIPTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type CLIPTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, QLIPVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type CLIPVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+        decoder_config = config.decoder_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+        self.text_model = QLIPTextTransformer(text_config)
+        self.vision_model = QLIPVisionTransformer(vision_config)
+        self.vision_decoder = QLIPVisionTransformerDecoder(decoder_config)
+
+        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=config.vision_projection_bias)
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=config.text_projection_bias)
+        self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`CLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use CLIP model'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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        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 [`CLIPVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use CLIP model'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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=QLIPOutput, config_class=QLIPConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QLIPOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPModel
+
+        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use CLIP model'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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(text_embeds)
+
+        last_hidden_state = vision_outputs[0]
+        recon = self.vision_decoder(last_hidden_state)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = clip_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return QLIPOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+            reconstructions=recon,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIP Text Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLIP_START_DOCSTRING,
+)
+class QLIPTextModelWithProjection(QLIPPreTrainedModel):
+    config_class = QLIPTextConfig
+
+    _no_split_modules = ["QLIPTextEmbeddings", "QLIPEncoderLayer"]
+
+    def __init__(self, config: QLIPTextConfig):
+        super().__init__(config)
+
+        self.text_model = QLIPTextTransformer(config)
+
+        self.text_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=QLIPTextModelOutput, config_class=QLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QLIPTextModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CLIPTextModelWithProjection
+
+        >>> model = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> text_embeds = outputs.text_embeds
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1]
+
+        text_embeds = self.text_projection(pooled_output)
+
+        if not return_dict:
+            outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return QLIPTextModelOutput(
+            text_embeds=text_embeds,
+            last_hidden_state=text_outputs.last_hidden_state,
+            hidden_states=text_outputs.hidden_states,
+            attentions=text_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLIP Vision Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLIP_START_DOCSTRING,
+)
+class QLIPVisionModelWithProjection(QLIPPreTrainedModel):
+    config_class = QLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: QLIPVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = QLIPVisionTransformer(config)
+
+        self.visual_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=QLIPVisionModelOutput, config_class=QLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QLIPVisionModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, CLIPVisionModelWithProjection
+
+        >>> model = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
+        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> 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)
+        >>> image_embeds = outputs.image_embeds
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+
+        image_embeds = self.visual_projection(pooled_output)
+
+        if not return_dict:
+            outputs = (image_embeds, vision_outputs[0]) + vision_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return QLIPVisionModelOutput(
+            image_embeds=image_embeds,
+            last_hidden_state=vision_outputs.last_hidden_state,
+            hidden_states=vision_outputs.hidden_states,
+            attentions=vision_outputs.attentions,
+        )

preprocessor_config.json

@@ -0,0 +1,19 @@
+{
+  "crop_size": 392,
+  "do_center_crop": true,
+  "do_normalize": true,
+  "do_resize": true,
+  "feature_extractor_type": "CLIPFeatureExtractor",
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "resample": 3,
+  "size": 392
+}

rope.py

@@ -0,0 +1,118 @@
+# Copyright (c) 2024, NVIDIA Corporation & Affiliates. All rights reserved. 
+# 
+# This work is made available under the Nvidia Source Code License-NC. 
+# To view a copy of this license, visit 
+# https://github.com/NVlabs/QLIP/blob/main/LICENSE
+
+# MIT License
+
+# Copyright (c) 2022 BAAI-Vision
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+from math import pi
+import torch
+from torch import nn
+from einops import rearrange, repeat
+import logging
+
+
+def broadcat(tensors, dim = -1):
+    num_tensors = len(tensors)
+    shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
+    assert len(shape_lens) == 1, 'tensors must all have the same number of dimensions'
+    shape_len = list(shape_lens)[0]
+    dim = (dim + shape_len) if dim < 0 else dim
+    dims = list(zip(*map(lambda t: list(t.shape), tensors)))
+    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+    assert all([*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]), 'invalid dimensions for broadcastable concatentation'
+    max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
+    expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
+    expanded_dims.insert(dim, (dim, dims[dim]))
+    expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
+    tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
+    return torch.cat(tensors, dim = dim)
+
+def rotate_half(x):
+    x = rearrange(x, '... (d r) -> ... d r', r = 2)
+    x1, x2 = x.unbind(dim = -1)
+    x = torch.stack((-x2, x1), dim = -1)
+    return rearrange(x, '... d r -> ... (d r)')
+
+
+class VisionRotaryEmbeddingFast(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs = None,
+        freqs_for = 'lang',
+        theta = 10000,
+        max_freq = 10,
+        num_freqs = 1,
+        patch_dropout = 0.
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None: ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs = torch.einsum('..., f -> ... f', t, freqs)
+        freqs = repeat(freqs, '... n -> ... (n r)', r = 2)
+        freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim = -1)
+
+        freqs_cos = freqs.cos().view(-1, freqs.shape[-1])
+        freqs_sin = freqs.sin().view(-1, freqs.shape[-1])
+
+        self.patch_dropout = patch_dropout
+
+        self.register_buffer("freqs_cos", freqs_cos)
+        self.register_buffer("freqs_sin", freqs_sin)
+
+        logging.info(f'Shape of rope freq: {self.freqs_cos.shape}')
+
+    def forward(self, t, patch_indices_keep=None):
+        if patch_indices_keep is not None:
+            batch = t.size()[0]
+            batch_indices = torch.arange(batch)
+            batch_indices = batch_indices[..., None]
+
+            freqs_cos = repeat(self.freqs_cos, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+            freqs_sin = repeat(self.freqs_sin, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+
+            freqs_cos = freqs_cos[batch_indices, patch_indices_keep]
+            freqs_cos = rearrange(freqs_cos, 'n i m j -> n m i j')
+            freqs_sin = freqs_sin[batch_indices, patch_indices_keep]
+            freqs_sin = rearrange(freqs_sin, 'n i m j -> n m i j')
+
+            return  t * freqs_cos + rotate_half(t) * freqs_sin
+
+        return  t * self.freqs_cos + rotate_half(t) * self.freqs_sin

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": {"content": "<|startoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "pad_token": "<|endoftext|>"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "<|startoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "pad_token": "<|endoftext|>", "add_prefix_space": false, "errors": "replace", "do_lower_case": true, "name_or_path": "openai/clip-vit-base-patch32", "model_max_length": 77, "special_tokens_map_file": "./special_tokens_map.json", "tokenizer_class": "CLIPTokenizer"}