Initial commit

.gitattributes

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

README.md

@@ -0,0 +1,202 @@
+---
+base_model: openbmb/MiniCPM-V-2_6
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.11.1

adapter_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:275ba4ef8d4ca33b1e83426f1ca156b900b0bcdf093e52cdc9d4da02a6ca006c
+size 886

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:675fa04d31250951b24651dcdd0d9952727b18b805ca1137f874d6e7af9f8ca4
+size 2140796552

added_tokens.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f67327aead2a3259804bdbb2a7b592c5238af4434f92b7952555f16531228804
+size 629

config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ca8152bcc0655f23002e12e9b910e4ced9793310014eea18c5b291e1b887ed2
+size 1365

configuration_minicpm.py

@@ -0,0 +1,100 @@
+# coding=utf-8
+""" MiniCPMV model configuration"""
+
+import os
+from typing import Union
+
+from transformers.utils import logging
+from transformers import Qwen2Config, PretrainedConfig
+from .modeling_navit_siglip import SiglipVisionConfig
+
+logger = logging.get_logger(__name__)
+
+
+class MiniCPMVSliceConfig(PretrainedConfig):
+    model_type = "minicpmv"
+
+    def __init__(
+        self,
+        patch_size=14,
+        max_slice_nums=9,
+        scale_resolution=448,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        self.max_slice_nums = max_slice_nums
+        self.scale_resolution = scale_resolution
+
+    @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)
+
+        if config_dict.get("model_type") == "minicpmv":
+            config_dict = config_dict["slice_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 MiniCPMVConfig(Qwen2Config):
+    model_type = "minicpmv"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    default_vision_config = {
+        "hidden_size": 1152,
+        "image_size": 980,
+        "intermediate_size": 4304,
+        "model_type": "siglip",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 27,
+        "patch_size": 14,
+    }
+
+    def __init__(
+        self,
+        use_cache=True,
+        query_num=64,
+        image_size=448,
+        drop_vision_last_layer=True,
+        batch_vision_input=True,
+        slice_config=None,
+        vision_config=None,
+        use_image_id=True,
+        vision_batch_size=16,
+        **kwargs,
+    ):
+        self.use_cache = use_cache
+        self.query_num = query_num
+        self.image_size = image_size
+        self.drop_vision_last_layer = drop_vision_last_layer
+        self.batch_vision_input = batch_vision_input
+        self.use_image_id = use_image_id
+        self.vision_batch_size = vision_batch_size
+
+        if slice_config is None:
+            self.slice_config = MiniCPMVSliceConfig(max_slice_nums=1)
+        else:
+            self.slice_config = MiniCPMVSliceConfig(**slice_config)
+        self.slice_mode = True
+
+        # same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit add tgt_sizes
+        if vision_config is None:
+            self.vision_config = SiglipVisionConfig(**self.default_vision_config)
+            logger.info("vision_config is None, using default vision config")
+        elif isinstance(vision_config, dict):
+            self.vision_config = SiglipVisionConfig(**vision_config)
+        elif isinstance(vision_config, SiglipVisionConfig):
+            self.vision_config = vision_config
+
+        self.patch_size = self.vision_config.patch_size
+
+        super().__init__(**kwargs)

generation_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f32c18467b4816e3e93ef405b8d247f3607c2dbf14d97031b90894739016237
+size 121

image_processing_minicpmv.py

@@ -0,0 +1,418 @@
+from typing import Optional, Union, Dict, Any, List
+
+import torch
+import math
+import PIL.Image
+import PIL.ImageSequence
+import numpy as np
+import PIL
+from PIL import Image
+
+from transformers.utils import TensorType, requires_backends, is_torch_dtype, is_torch_device
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers import AutoImageProcessor
+from transformers.image_transforms import to_channel_dimension_format
+from transformers.image_utils import (
+    ImageInput, 
+    make_list_of_images, 
+    valid_images, 
+    is_torch_tensor, 
+    is_batched,
+    to_numpy_array, 
+    infer_channel_dimension_format,
+    ChannelDimension
+)
+
+
+def recursive_converter(converter, value):
+    if isinstance(value, list):
+        new_value = []
+        for v in value:
+            new_value += [recursive_converter(converter, v)]
+        return new_value
+    else:
+        return converter(value)
+
+
+class MiniCPMVBatchFeature(BatchFeature):
+    r"""
+    Extend from BatchFeature for supporting various image size
+    """
+    def __init__(self, data: Optional[Dict[str, Any]] = None, tensor_type: Union[None, str, TensorType] = None):
+        super().__init__(data)
+        self.convert_to_tensors(tensor_type=tensor_type)
+
+    def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = None):
+        if tensor_type is None:
+            return self
+        
+        is_tensor, as_tensor = self._get_is_as_tensor_fns(tensor_type)
+
+        def converter(value):
+            try:
+                if not is_tensor(value):
+                    tensor = as_tensor(value)
+                    return tensor
+            except:  # noqa E722
+                if key == "overflowing_values":
+                    raise ValueError("Unable to create tensor returning overflowing values of different lengths. ")
+                raise ValueError(
+                    "Unable to create tensor, you should probably activate padding "
+                    "with 'padding=True' to have batched tensors with the same length."
+                )
+
+
+        for key, value in self.items():
+            self[key] = recursive_converter(converter, value)
+        return self
+            
+    def to(self, *args, **kwargs) -> "MiniCPMVBatchFeature":
+        requires_backends(self, ["torch"])
+        import torch
+
+        def cast_tensor(v):
+            # check if v is a floating point
+            if torch.is_floating_point(v):
+                # cast and send to device
+                return v.to(*args, **kwargs)
+            elif device is not None:
+                return v.to(device=device)
+            else:
+                return v
+
+        new_data = {}
+        device = kwargs.get("device")
+        # Check if the args are a device or a dtype
+        if device is None and len(args) > 0:
+            # device should be always the first argument
+            arg = args[0]
+            if is_torch_dtype(arg):
+                # The first argument is a dtype
+                pass
+            elif isinstance(arg, str) or is_torch_device(arg) or isinstance(arg, int):
+                device = arg
+            else:
+                # it's something else
+                raise ValueError(f"Attempting to cast a BatchFeature to type {str(arg)}. This is not supported.")
+        # We cast only floating point tensors to avoid issues with tokenizers casting `LongTensor` to `FloatTensor`
+        for k, v in self.items():
+            new_data[k] = recursive_converter(cast_tensor, v)
+        self.data = new_data
+        return self
+
+
+class MiniCPMVImageProcessor(BaseImageProcessor):
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+            self, 
+            max_slice_nums=9,
+            scale_resolution=448,
+            patch_size=14,
+            **kwargs):
+        super().__init__(**kwargs)
+        self.max_slice_nums = max_slice_nums
+        self.scale_resolution = scale_resolution
+        self.patch_size = patch_size
+        self.use_image_id = kwargs.pop("use_image_id", False)
+        self.image_feature_size = kwargs.pop("image_feature_size", 64)
+        self.im_start_token = kwargs.pop("im_start", "<image>")
+        self.im_end_token = kwargs.pop("im_end", "</image>")
+        self.slice_start_token = kwargs.pop("slice_start", "<slice>")
+        self.slice_end_token = kwargs.pop("slice_end", "</slice>")
+        self.unk_token = kwargs.pop("unk", "<unk>")
+        self.im_id_start = kwargs.pop("im_id_start", "<image_id>")
+        self.im_id_end = kwargs.pop("im_id_end", "</image_id>")
+        self.slice_mode = kwargs.pop("slice_mode", True)
+        self.mean = np.array(kwargs.pop("norm_mean", [0.5, 0.5, 0.5]))
+        self.std = np.array(kwargs.pop("norm_std", [0.5, 0.5, 0.5]))
+        self.version = kwargs.pop("version", 2.0)
+
+    def ensure_divide(self, length, patch_size):
+        return max(round(length / patch_size) * patch_size, patch_size)
+
+    def find_best_resize(self,
+                         original_size,
+                         scale_resolution,
+                         patch_size,
+                         allow_upscale=False):
+        width, height = original_size
+        if (width * height >
+                scale_resolution * scale_resolution) or allow_upscale:
+            r = width / height
+            height = int(scale_resolution / math.sqrt(r))
+            width = int(height * r)
+        best_width = self.ensure_divide(width, patch_size)
+        best_height = self.ensure_divide(height, patch_size)
+        return (best_width, best_height)
+
+    def get_refine_size(self,
+                        original_size,
+                        grid,
+                        scale_resolution,
+                        patch_size,
+                        allow_upscale=False):
+        width, height = original_size
+        grid_x, grid_y = grid
+
+        refine_width = self.ensure_divide(width, grid_x)
+        refine_height = self.ensure_divide(height, grid_y)
+
+        grid_width = refine_width / grid_x
+        grid_height = refine_height / grid_y
+
+        best_grid_size = self.find_best_resize((grid_width, grid_height),
+                                               scale_resolution,
+                                               patch_size,
+                                               allow_upscale=allow_upscale)
+        refine_size = (best_grid_size[0] * grid_x, best_grid_size[1] * grid_y)
+        return refine_size
+
+    def split_to_patches(self, image, grid):
+        patches = []
+        width, height = image.size
+        grid_x = int(width / grid[0])
+        grid_y = int(height / grid[1])
+        for i in range(0, height, grid_y):
+            images = []
+            for j in range(0, width, grid_x):
+                box = (j, i, j + grid_x, i + grid_y)
+                patch = image.crop(box)
+                images.append(patch)
+            patches.append(images)
+        return patches
+
+    def slice_image(
+        self, image, max_slice_nums=9, scale_resolution=448, patch_size=14, never_split=False
+    ):
+        original_size = image.size
+        source_image = None
+        best_grid = self.get_sliced_grid(original_size, max_slice_nums, never_split)
+        patches = []
+
+        if best_grid is None:
+            # dont need to slice, upsample
+            best_size = self.find_best_resize(
+                original_size, scale_resolution, patch_size, allow_upscale=True
+            )
+            source_image = image.resize(best_size, resample=Image.Resampling.BICUBIC)
+        else:
+            # source image, down-sampling and ensure divided by patch_size
+            best_resize = self.find_best_resize(original_size, scale_resolution, patch_size)
+            source_image = image.copy().resize(best_resize, resample=Image.Resampling.BICUBIC)
+            refine_size = self.get_refine_size(
+                original_size, best_grid, scale_resolution, patch_size, allow_upscale=True
+            )
+            refine_image = image.resize(refine_size, resample=Image.Resampling.BICUBIC)
+            patches = self.split_to_patches(refine_image, best_grid)
+
+        return source_image, patches, best_grid
+
+    def get_grid_placeholder(self, grid):
+        if grid is None:
+            return ""
+        slice_image_placeholder = (
+            self.slice_start_token 
+            + self.unk_token * self.image_feature_size
+            + self.slice_end_token
+        )
+
+        cols = grid[0]
+        rows = grid[1]
+        slices = []
+        for i in range(rows):
+            lines = []
+            for j in range(cols):
+                lines.append(slice_image_placeholder)
+            slices.append("".join(lines))
+            
+        slice_placeholder = "\n".join(slices)
+        return slice_placeholder
+
+    def get_image_id_placeholder(self, idx=0):
+        return f"{self.im_id_start}{idx}{self.im_id_end}"
+    
+    def get_sliced_images(self, image, max_slice_nums=None):
+        slice_images = []
+
+        if not self.slice_mode:
+            return [image]
+
+        max_slice_nums = self.max_slice_nums if max_slice_nums is None else int(max_slice_nums)
+        assert max_slice_nums > 0 
+        source_image, patches, sliced_grid = self.slice_image(
+            image,
+            max_slice_nums,  # default: 9
+            self.scale_resolution,  # default: 448
+            self.patch_size  # default: 14
+        )
+
+        slice_images.append(source_image)
+        if len(patches) > 0:
+            for i in range(len(patches)):
+                for j in range(len(patches[0])):
+                    slice_images.append(patches[i][j])
+        return slice_images
+
+    def get_sliced_grid(self, image_size, max_slice_nums, nerver_split=False):
+        original_width, original_height = image_size
+        log_ratio = math.log(original_width / original_height)
+        ratio = original_width * original_height / (self.scale_resolution * self.scale_resolution)
+        multiple = min(math.ceil(ratio), max_slice_nums)
+        if multiple <= 1 or nerver_split:
+            return None
+        candidate_split_grids_nums = []
+        for i in [multiple - 1, multiple, multiple + 1]:
+            if i == 1 or i > max_slice_nums:
+                continue
+            candidate_split_grids_nums.append(i)
+        
+        candidate_grids = []
+        for split_grids_nums in candidate_split_grids_nums:
+            m = 1
+            while m <= split_grids_nums:
+                if split_grids_nums % m == 0:
+                    candidate_grids.append([m, split_grids_nums // m])
+                m += 1
+
+        best_grid = [1, 1]
+        min_error = float("inf")
+        for grid in candidate_grids:
+            error = abs(log_ratio - math.log(grid[0] / grid[1]))
+            if error < min_error:
+                best_grid = grid
+                min_error = error
+        
+        return best_grid
+    
+    def get_slice_image_placeholder(self, image_size, image_idx=0, max_slice_nums=None, use_image_id=None):
+        max_slice_nums = self.max_slice_nums if max_slice_nums is None else int(max_slice_nums)
+        assert max_slice_nums > 0        
+        grid = self.get_sliced_grid(image_size=image_size, max_slice_nums=max_slice_nums)
+
+        image_placeholder = (
+            self.im_start_token 
+            + self.unk_token * self.image_feature_size 
+            + self.im_end_token
+        )
+        use_image_id = self.use_image_id if use_image_id is None else bool(use_image_id)
+        if use_image_id:
+            final_placeholder = self.get_image_id_placeholder(image_idx) + image_placeholder
+        else:
+            final_placeholder = image_placeholder
+
+        if self.slice_mode:
+            final_placeholder = final_placeholder + self.get_grid_placeholder(grid=grid)
+        return final_placeholder
+        
+    def to_pil_image(self, image, rescale=None) -> PIL.Image.Image:
+        """
+        Converts `image` to a PIL Image. Optionally rescales it and puts the channel dimension back as the last axis if
+        needed.
+
+        Args:
+            image (`PIL.Image.Image` or `numpy.ndarray` or `torch.Tensor`):
+                The image to convert to the PIL Image format.
+            rescale (`bool`, *optional*):
+                Whether or not to apply the scaling factor (to make pixel values integers between 0 and 255). Will
+                default to `True` if the image type is a floating type, `False` otherwise.
+        """
+        if isinstance(image, PIL.Image.Image):
+            return image
+        if is_torch_tensor(image):
+            image = image.numpy()
+
+        if isinstance(image, np.ndarray):
+            if rescale is None:
+                # rescale default to the array being of floating type.
+                rescale = isinstance(image.flat[0], np.floating)
+            # If the channel as been moved to first dim, we put it back at the end.
+            if image.ndim == 3 and image.shape[0] in [1, 3]:
+                image = image.transpose(1, 2, 0)
+            if rescale:
+                image = image * 255
+            image = image.astype(np.uint8)
+            return PIL.Image.fromarray(image)
+        return image
+
+    def reshape_by_patch(self, image):
+        """
+        :param image: shape [3, H, W]
+        :param patch_size:
+        :return: [3, patch_size, HW/patch_size]
+        """
+        image = torch.from_numpy(image)
+        patch_size = self.patch_size
+        patches = torch.nn.functional.unfold(
+            image,
+            (patch_size, patch_size),
+            stride=(patch_size, patch_size)
+        )
+
+        patches = patches.reshape(image.size(0), patch_size, patch_size, -1)
+        patches = patches.permute(0, 1, 3, 2).reshape(image.size(0), patch_size, -1)
+        return patches.numpy()
+
+    def preprocess(
+            self, 
+            images: Union[Image.Image, List[Image.Image], List[List[Image.Image]]],
+            do_pad: Optional[bool] = True, # TODO: add pad for MiniCPM-Llama3-V-2_5
+            max_slice_nums: int = None,
+            return_tensors: Optional[Union[str, TensorType]] = None,
+            **kwargs
+        ) -> MiniCPMVBatchFeature:
+        if isinstance(images, Image.Image):
+            images_list = [[images]]
+        elif isinstance(images[0], Image.Image):
+            images_list = [images]
+        else:
+            images_list = images
+
+        new_images_list = []
+        image_sizes_list = []
+        tgt_sizes_list = []
+        
+        for _images in images_list:
+            if _images is None or len(_images) == 0:
+                new_images_list.append([])
+                image_sizes_list.append([])
+                tgt_sizes_list.append([])
+                continue             
+            if not valid_images(_images):
+                raise ValueError(
+                    "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                    "torch.Tensor, tf.Tensor or jax.ndarray."
+                )
+            
+            _images = [self.to_pil_image(image).convert("RGB") for image in _images]
+            input_data_format = infer_channel_dimension_format(np.array(_images[0]))
+
+            new_images = []
+            image_sizes = [image.size for image in _images]
+            tgt_sizes = []
+            for image in _images:
+                image_patches = self.get_sliced_images(image, max_slice_nums)
+                image_patches = [to_numpy_array(image).astype(np.float32) / 255 for image in image_patches]
+                image_patches = [
+                    self.normalize(image=image, mean=self.mean, std=self.std, input_data_format=input_data_format)
+                        for image in image_patches
+                ]
+                image_patches = [
+                    to_channel_dimension_format(image, ChannelDimension.FIRST, input_channel_dim=input_data_format) 
+                        for image in image_patches
+                ]
+                for slice_image in image_patches:
+                    new_images.append(self.reshape_by_patch(slice_image))
+                    tgt_sizes.append(np.array((slice_image.shape[1] // self.patch_size, slice_image.shape[2] // self.patch_size)))
+
+            if tgt_sizes:
+                tgt_sizes = np.vstack(tgt_sizes)
+            
+            new_images_list.append(new_images)
+            image_sizes_list.append(image_sizes)
+            tgt_sizes_list.append(tgt_sizes)
+        return MiniCPMVBatchFeature(
+            data={"pixel_values": new_images_list, "image_sizes": image_sizes_list, "tgt_sizes": tgt_sizes_list}, tensor_type=return_tensors
+        )
+
+AutoImageProcessor.register("MiniCPMVImageProcessor", MiniCPMVImageProcessor)

latest

@@ -0,0 +1 @@
+global_step100000

modeling_minicpmv.py

@@ -0,0 +1,403 @@
+import math
+from typing import List, Optional
+import json
+import torch
+import torchvision
+
+from threading import Thread
+from copy import deepcopy
+from PIL import Image
+from transformers import AutoProcessor, Qwen2PreTrainedModel, Qwen2ForCausalLM, TextIteratorStreamer
+
+from .configuration_minicpm import MiniCPMVConfig
+from .modeling_navit_siglip import SiglipVisionTransformer
+from .resampler import Resampler
+
+
+
+class MiniCPMVPreTrainedModel(Qwen2PreTrainedModel):
+    config_class = MiniCPMVConfig
+
+
+class MiniCPMV(MiniCPMVPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.llm = Qwen2ForCausalLM(config)
+        self.vpm = self.init_vision_module()
+        self.vision_dim = self.vpm.embed_dim
+        self.embed_dim = self.llm.config.hidden_size
+        self.resampler = self.init_resampler(self.embed_dim, self.vision_dim)
+        self.processor = None
+
+        self.terminators = ['<|im_end|>', '<|endoftext|>']
+
+    def init_vision_module(self):
+        # same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit add tgt_sizes
+        if self.config._attn_implementation == 'flash_attention_2':
+            self.config.vision_config._attn_implementation = 'flash_attention_2'
+        else:
+            # not suport sdpa
+            self.config.vision_config._attn_implementation = 'eager'
+        model = SiglipVisionTransformer(self.config.vision_config)
+        if self.config.drop_vision_last_layer:
+            model.encoder.layers = model.encoder.layers[:-1]
+
+        setattr(model, 'embed_dim', model.embeddings.embed_dim)
+        setattr(model, 'patch_size', model.embeddings.patch_size)
+
+        return model
+
+    def init_resampler(self, embed_dim, vision_dim):
+        return Resampler(
+            num_queries=self.config.query_num,
+            embed_dim=embed_dim,
+            num_heads=embed_dim // 128,
+            kv_dim=vision_dim,
+            adaptive=True
+        )
+
+    def get_input_embeddings(self):
+        return self.llm.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.llm.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.llm.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.llm.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.llm = decoder
+
+    def get_decoder(self):
+        return self.llm
+
+    def get_vllm_embedding(self, data):
+        if 'vision_hidden_states' not in data:
+            dtype = self.llm.model.embed_tokens.weight.dtype
+            device = self.llm.model.embed_tokens.weight.device
+            tgt_sizes = data['tgt_sizes']
+            pixel_values_list = data['pixel_values']
+            vision_hidden_states = []
+            all_pixel_values = []
+            img_cnt = []
+            for pixel_values in pixel_values_list:
+                img_cnt.append(len(pixel_values))
+                all_pixel_values.extend([i.flatten(end_dim=1).permute(1, 0) for i in pixel_values])
+
+            # exist image
+            if all_pixel_values:
+                tgt_sizes = [tgt_size for tgt_size in tgt_sizes if isinstance(tgt_size, torch.Tensor)]
+                tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)
+
+                max_patches = torch.max(tgt_sizes[:, 0] * tgt_sizes[:, 1])
+
+                all_pixel_values = torch.nn.utils.rnn.pad_sequence(all_pixel_values, batch_first=True,
+                                                                   padding_value=0.0)
+                B, L, _ = all_pixel_values.shape
+                all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
+
+                patch_attn_mask = torch.zeros((B, 1, max_patches), dtype=torch.bool, device=device)
+                for i in range(B):
+                    patch_attn_mask[i, 0, :tgt_sizes[i][0] * tgt_sizes[i][1]] = True
+
+                vision_batch_size = self.config.vision_batch_size
+                all_pixel_values = all_pixel_values.type(dtype)
+                if B > vision_batch_size:
+                    hs = []
+                    for i in range(0, B, vision_batch_size):
+                        start_idx = i
+                        end_idx = i + vision_batch_size
+                        tmp_hs = self.vpm(all_pixel_values[start_idx:end_idx], patch_attention_mask=patch_attn_mask[start_idx:end_idx], tgt_sizes=tgt_sizes[start_idx:end_idx]).last_hidden_state
+                        hs.append(tmp_hs)
+                    vision_embedding = torch.cat(hs, dim=0)
+                else:
+                    vision_embedding = self.vpm(all_pixel_values, patch_attention_mask=patch_attn_mask, tgt_sizes=tgt_sizes).last_hidden_state
+                vision_embedding = self.resampler(vision_embedding, tgt_sizes)
+
+                start = 0
+                for pixel_values in pixel_values_list:
+                    img_cnt = len(pixel_values)
+                    if img_cnt > 0:
+                        vision_hidden_states.append(vision_embedding[start: start + img_cnt])
+                        start += img_cnt
+                    else:
+                        vision_hidden_states.append([])
+            else: # no image
+                if self.training:
+                    dummy_image = torch.zeros(
+                        (1, 3, 224, 224),
+                        device=device, dtype=dtype
+                    )
+                    tgt_sizes = torch.Tensor([[(224 // self.config.patch_size), math.ceil(224 / self.config.patch_size)]]).type(torch.int32)
+                    dummy_feature = self.resampler(self.vpm(dummy_image).last_hidden_state, tgt_sizes)
+                else:
+                    dummy_feature = []
+                for _ in range(len(pixel_values_list)):
+                    vision_hidden_states.append(dummy_feature)
+
+        else:
+            vision_hidden_states = data['vision_hidden_states']
+
+        if hasattr(self.llm.config, 'scale_emb'):
+            vllm_embedding = self.llm.model.embed_tokens(data['input_ids']) * self.llm.config.scale_emb
+        else:
+            vllm_embedding = self.llm.model.embed_tokens(data['input_ids'])
+
+        vision_hidden_states = [i.type(vllm_embedding.dtype) if isinstance(
+            i, torch.Tensor) else i for i in vision_hidden_states]
+
+        bs = len(data['input_ids'])
+        for i in range(bs):
+            cur_vs_hs = vision_hidden_states[i]
+            if len(cur_vs_hs) > 0:
+                cur_vllm_emb = vllm_embedding[i]
+                cur_image_bound = data['image_bound'][i]
+                if len(cur_image_bound) > 0:
+                    image_indices = torch.stack(
+                        [torch.arange(r[0], r[1], dtype=torch.long) for r in cur_image_bound]
+                    ).to(vllm_embedding.device)
+
+                    cur_vllm_emb.scatter_(0, image_indices.view(-1, 1).repeat(1, cur_vllm_emb.shape[-1]),
+                                          cur_vs_hs.view(-1, cur_vs_hs.shape[-1]))
+                elif self.training:
+                    cur_vllm_emb += cur_vs_hs[0].mean() * 0
+
+        return vllm_embedding, vision_hidden_states
+
+    def forward(self, data, **kwargs):
+        vllm_embedding, vision_hidden_states = self.get_vllm_embedding(data)
+        position_ids = data["position_ids"]
+        if position_ids.dtype != torch.int64:
+            position_ids = position_ids.long()
+
+        return self.llm(
+            input_ids=None,
+            position_ids=position_ids,
+            inputs_embeds=vllm_embedding,
+            **kwargs
+        )
+    
+    def _decode(self, inputs_embeds, tokenizer, attention_mask, decode_text=False, **kwargs):
+        terminators = [tokenizer.convert_tokens_to_ids(i) for i in self.terminators]
+        output = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            pad_token_id=0,
+            eos_token_id=terminators,
+            attention_mask=attention_mask,
+            **kwargs
+        )
+        if decode_text:
+            return self._decode_text(output, tokenizer)
+        return output
+
+    def _decode_stream(self, inputs_embeds, tokenizer, **kwargs):
+        terminators = [tokenizer.convert_tokens_to_ids(i) for i in self.terminators]
+        streamer = TextIteratorStreamer(tokenizer=tokenizer)
+        generation_kwargs = {
+            'inputs_embeds': inputs_embeds,
+            'pad_token_id': 0,
+            'eos_token_id': terminators,
+            'streamer': streamer
+        }
+        generation_kwargs.update(kwargs)
+
+        thread = Thread(target=self.llm.generate, kwargs=generation_kwargs)
+        thread.start()
+    
+        return streamer
+
+    def _decode_text(self, result_ids, tokenizer):
+        terminators = [tokenizer.convert_tokens_to_ids(i) for i in self.terminators]
+        result_text = []
+        for result in result_ids:
+            result = result[result != 0]
+            if result[0] == tokenizer.bos_id:
+                result = result[1:]
+            if result[-1] in terminators:
+                result = result[:-1]
+            result_text.append(tokenizer.decode(result).strip())
+        return result_text
+
+    def generate(
+        self,
+        input_ids=None,
+        pixel_values=None,
+        tgt_sizes=None,
+        image_bound=None,
+        attention_mask=None,
+        tokenizer=None,
+        vision_hidden_states=None,
+        return_vision_hidden_states=False,
+        stream=False,
+        decode_text=False,
+        **kwargs
+    ):
+        assert input_ids is not None
+        assert len(input_ids) == len(pixel_values)
+
+        model_inputs = {
+            "input_ids": input_ids,
+            "image_bound": image_bound,
+        }
+
+        if vision_hidden_states is None:
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs['tgt_sizes'] = tgt_sizes
+        else:
+            model_inputs["vision_hidden_states"] = vision_hidden_states
+
+        with torch.inference_mode():
+            (
+                model_inputs["inputs_embeds"],
+                vision_hidden_states,
+            ) = self.get_vllm_embedding(model_inputs)
+
+            if stream:
+                result = self._decode_stream(model_inputs["inputs_embeds"], tokenizer, **kwargs)
+            else:
+                result = self._decode(model_inputs["inputs_embeds"], tokenizer, attention_mask, decode_text=decode_text, **kwargs)
+
+        if return_vision_hidden_states:
+            return result, vision_hidden_states
+        
+        return result
+
+    def chat(
+        self,
+        image,
+        msgs,
+        tokenizer,
+        processor=None,
+        vision_hidden_states=None,
+        max_new_tokens=2048,
+        min_new_tokens=0,
+        sampling=True,
+        max_inp_length=8192,
+        system_prompt='',
+        stream=False,
+        max_slice_nums=None,
+        use_image_id=None,
+        **kwargs
+    ):
+        if isinstance(msgs[0], list):
+            batched = True
+        else:
+            batched = False
+        msgs_list = msgs
+        images_list = image
+        
+        if batched is False:
+            images_list, msgs_list = [images_list], [msgs_list]
+        else:
+            assert images_list is None, "Please integrate image to msgs when using batch inference."
+            images_list = [None] * len(msgs_list)
+        assert len(images_list) == len(msgs_list), "The batch dim of images_list and msgs_list should be the same."
+
+        if processor is None:
+            if self.processor is None:
+                self.processor = AutoProcessor.from_pretrained(self.config._name_or_path, trust_remote_code=True)
+            processor = self.processor
+        
+        assert self.config.query_num == processor.image_processor.image_feature_size, "These two values should be the same. Check `config.json` and `preprocessor_config.json`."
+        assert self.config.patch_size == processor.image_processor.patch_size, "These two values should be the same. Check `config.json` and `preprocessor_config.json`."
+        assert self.config.use_image_id == processor.image_processor.use_image_id, "These two values should be the same. Check `config.json` and `preprocessor_config.json`."
+        assert self.config.slice_config.max_slice_nums == processor.image_processor.max_slice_nums, "These two values should be the same. Check `config.json` and `preprocessor_config.json`."
+        assert self.config.slice_mode == processor.image_processor.slice_mode, "These two values should be the same. Check `config.json` and `preprocessor_config.json`."
+
+        prompts_lists = []
+        input_images_lists = []
+        for image, msgs in zip(images_list, msgs_list):
+            if isinstance(msgs, str):
+                msgs = json.loads(msgs)
+            copy_msgs = deepcopy(msgs)
+
+            assert len(msgs) > 0, "msgs is empty"
+            assert sampling or not stream, "if use stream mode, make sure sampling=True"
+
+            if image is not None and isinstance(copy_msgs[0]["content"], str):
+                copy_msgs[0]["content"] = [image, copy_msgs[0]["content"]]
+
+            images = []
+            for i, msg in enumerate(copy_msgs):
+                role = msg["role"]
+                content = msg["content"]
+                assert role in ["user", "assistant"]
+                if i == 0:
+                    assert role == "user", "The role of first msg should be user"
+                if isinstance(content, str):
+                    content = [content]
+                cur_msgs = []
+                for c in content:
+                    if isinstance(c, Image.Image):
+                        images.append(c)
+                        cur_msgs.append("(<image>./</image>)")
+                    elif isinstance(c, str):
+                        cur_msgs.append(c)
+                msg["content"] = "\n".join(cur_msgs)
+
+            if system_prompt:
+                sys_msg = {'role': 'system', 'content': system_prompt}
+                copy_msgs = [sys_msg] + copy_msgs        
+
+            prompts_lists.append(processor.tokenizer.apply_chat_template(copy_msgs, tokenize=False, add_generation_prompt=True))
+            input_images_lists.append(images)
+
+        inputs = processor(
+            prompts_lists, 
+            input_images_lists, 
+            max_slice_nums=max_slice_nums,
+            use_image_id=use_image_id,
+            return_tensors="pt", 
+            max_length=max_inp_length
+        ).to(self.device)
+
+        if sampling:
+            generation_config = {
+                "top_p": 0.8,
+                "top_k": 100,
+                "temperature": 0.7,
+                "do_sample": True,
+                "repetition_penalty": 1.05
+            }
+        else:
+            generation_config = {
+                "num_beams": 3,
+                "repetition_penalty": 1.2,
+            }
+            
+        if min_new_tokens > 0:
+            generation_config['min_new_tokens'] = min_new_tokens
+
+        generation_config.update(
+            (k, kwargs[k]) for k in generation_config.keys() & kwargs.keys()
+        )
+
+        inputs.pop("image_sizes")
+        with torch.inference_mode():
+            res = self.generate(
+                **inputs,
+                tokenizer=tokenizer,
+                max_new_tokens=max_new_tokens,
+                vision_hidden_states=vision_hidden_states,
+                stream=stream,
+                decode_text=True,
+                **generation_config
+            )
+        
+        if stream:
+            def stream_gen():
+                for text in res:
+                    for term in self.terminators:
+                        text = text.replace(term, '')
+                    yield text
+            return stream_gen()
+
+        else:
+            if batched:
+                answer = res
+            else:
+                answer = res[0]
+            return answer

modeling_navit_siglip.py

@@ -0,0 +1,937 @@
+# coding=utf-8
+# Copyright 2024 Google AI 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 Siglip model. """
+# Copied from  HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes
+
+
+import os
+import math
+import warnings
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class SiglipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip 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 Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    Example:
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @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 SiglipConfig
+        if config_dict.get("model_type") == "siglip":
+            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)
+        
+
+_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
+
+SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/siglip-base-patch16-224",
+    # See all SigLIP models at https://huggingface.co/models?filter=siglip
+]
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    if tensor.dtype in [torch.float16, torch.bfloat16]:
+        # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
+        og_dtype = tensor.dtype
+        tensor = tensor.to(torch.float32)
+        tensor.erfinv_()
+        tensor = tensor.to(og_dtype)
+    else:
+        tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    if tensor.dtype == torch.float16:
+        # The `clamp_` op is not (yet?) defined in float16+cpu
+        tensor = tensor.to(torch.float32)
+        tensor.clamp_(min=a, max=b)
+        tensor = tensor.to(torch.float16)
+    else:
+        tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsquently scaled and shifted by the mean and std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
+class SiglipVisionModelOutput(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
+
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.BoolTensor, tgt_sizes: Optional[torch.IntTensor]=None) -> torch.Tensor:
+        batch_size = pixel_values.size(0)
+
+        patch_embeds = self.patch_embedding(pixel_values)
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+
+        max_im_h, max_im_w = pixel_values.size(2), pixel_values.size(3)
+        max_nb_patches_h, max_nb_patches_w = max_im_h // self.patch_size, max_im_w // self.patch_size
+        boundaries = torch.arange(1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side)
+        position_ids = torch.full(
+            size=(
+                batch_size,
+                max_nb_patches_h * max_nb_patches_w,
+            ),
+            fill_value=0,
+        )
+
+        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
+            if tgt_sizes is not None:
+                nb_patches_h = tgt_sizes[batch_idx][0]
+                nb_patches_w = tgt_sizes[batch_idx][1]
+            else:
+                nb_patches_h = p_attn_mask[:, 0].sum()
+                nb_patches_w = p_attn_mask[0].sum()
+
+            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
+            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
+
+            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
+            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)
+
+            pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side + bucket_coords_w).flatten()
+            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
+
+        position_ids = position_ids.to(self.position_embedding.weight.device)
+
+        embeddings = embeddings + self.position_embedding(position_ids)
+        return embeddings
+
+
+class SiglipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        k_v_seq_len = key_states.shape[-2]
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
+
+        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class SiglipFlashAttention2(SiglipAttention):
+    """
+    Llama flash attention module. This module inherits from `LlamaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.is_causal = False  # Hack to make sure we don't use a causal mask
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        # cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        # if past_key_value is not None:
+        #     cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+        #     key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                "The input hidden states seems to be silently casted in float32, this might be related to the fact"
+                " you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights
+
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+
+        # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+        causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
+class SiglipEncoderLayer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.self_attn = (
+            SiglipAttention(config)
+            if not self._use_flash_attention_2
+            else SiglipFlashAttention2(config)
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(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,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class SiglipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SiglipVisionConfig
+    base_model_prefix = "siglip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+
+        if isinstance(module, SiglipVisionEmbeddings):
+            width = self.config.hidden_size
+            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
+        elif isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, SiglipAttention):
+            nn.init.normal_(module.q_proj.weight)
+            nn.init.normal_(module.k_proj.weight)
+            nn.init.normal_(module.v_proj.weight)
+            nn.init.normal_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, SiglipMLP):
+            nn.init.normal_(module.fc1.weight)
+            nn.init.normal_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+SIGLIP_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 ([`SiglipVisionConfig`]): 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.
+"""
+
+
+SIGLIP_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.
+"""
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
+class SiglipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`SiglipEncoderLayer`].
+    Args:
+        config: SiglipConfig
+    """
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    # Ignore copy
+    def forward(
+        self,
+        inputs_embeds,
+        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)
+            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 encoder_layer in 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,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        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
+        )
+
+@add_start_docstrings(
+    """The vision model from SigLIP without any head or projection on top.""",
+    SIGLIP_START_DOCSTRING
+)
+class SiglipVisionTransformer(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "pixel_values"
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SiglipVisionEmbeddings(config)
+        self.encoder = SiglipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipVisionConfig)
+    def forward(
+        self,
+        pixel_values,
+        patch_attention_mask: Optional[torch.BoolTensor] = None,
+        tgt_sizes: Optional[torch.IntTensor] = 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
+
+        batch_size = pixel_values.size(0)
+        if patch_attention_mask is None:
+            patch_attention_mask = torch.ones(
+                size=(
+                    batch_size,
+                    pixel_values.size(2) // self.config.patch_size,
+                    pixel_values.size(3) // self.config.patch_size,
+                ),
+                dtype=torch.bool,
+                device=pixel_values.device,
+            )
+
+        hidden_states = self.embeddings(pixel_values=pixel_values, patch_attention_mask=patch_attention_mask, tgt_sizes=tgt_sizes)
+
+        patch_attention_mask = patch_attention_mask.view(batch_size, -1)
+        # The call to `_upad_input` in `_flash_attention_forward` is expensive
+        # So when the `patch_attention_mask` is full of 1s (i.e. attending to the whole sequence),
+        # avoiding passing the attention_mask, which is equivalent to attending to the full sequence
+        if not torch.any(~patch_attention_mask):
+            attention_mask=None
+        else:
+            attention_mask = (
+                _prepare_4d_attention_mask(patch_attention_mask, hidden_states.dtype)
+                if not self._use_flash_attention_2
+                else patch_attention_mask
+            )
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=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.post_layernorm(last_hidden_state)
+
+        if not return_dict:
+            return (last_hidden_state, None) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=None,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:542dce431c772b61bed842c09e490e42c88bfe671f0a2cdb04a672b185e95107
+size 2141086636

preprocessor_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e11ecbfde6e5be5ee1b4050f878edc5737043b901a51b6b03b5b838c75003c0
+size 715

processing_minicpmv.py

@@ -0,0 +1,937 @@
+# coding=utf-8
+# Copyright 2024 Google AI 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 Siglip model. """
+# Copied from  HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes
+
+
+import os
+import math
+import warnings
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class SiglipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip 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 Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    Example:
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @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 SiglipConfig
+        if config_dict.get("model_type") == "siglip":
+            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)
+        
+
+_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
+
+SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/siglip-base-patch16-224",
+    # See all SigLIP models at https://huggingface.co/models?filter=siglip
+]
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    if tensor.dtype in [torch.float16, torch.bfloat16]:
+        # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
+        og_dtype = tensor.dtype
+        tensor = tensor.to(torch.float32)
+        tensor.erfinv_()
+        tensor = tensor.to(og_dtype)
+    else:
+        tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    if tensor.dtype == torch.float16:
+        # The `clamp_` op is not (yet?) defined in float16+cpu
+        tensor = tensor.to(torch.float32)
+        tensor.clamp_(min=a, max=b)
+        tensor = tensor.to(torch.float16)
+    else:
+        tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsquently scaled and shifted by the mean and std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
+class SiglipVisionModelOutput(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
+
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.BoolTensor, tgt_sizes: Optional[torch.IntTensor]=None) -> torch.Tensor:
+        batch_size = pixel_values.size(0)
+
+        patch_embeds = self.patch_embedding(pixel_values)
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+
+        max_im_h, max_im_w = pixel_values.size(2), pixel_values.size(3)
+        max_nb_patches_h, max_nb_patches_w = max_im_h // self.patch_size, max_im_w // self.patch_size
+        boundaries = torch.arange(1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side)
+        position_ids = torch.full(
+            size=(
+                batch_size,
+                max_nb_patches_h * max_nb_patches_w,
+            ),
+            fill_value=0,
+        )
+
+        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
+            if tgt_sizes is not None:
+                nb_patches_h = tgt_sizes[batch_idx][0]
+                nb_patches_w = tgt_sizes[batch_idx][1]
+            else:
+                nb_patches_h = p_attn_mask[:, 0].sum()
+                nb_patches_w = p_attn_mask[0].sum()
+
+            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
+            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
+
+            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
+            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)
+
+            pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side + bucket_coords_w).flatten()
+            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
+
+        position_ids = position_ids.to(self.position_embedding.weight.device)
+
+        embeddings = embeddings + self.position_embedding(position_ids)
+        return embeddings
+
+
+class SiglipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        k_v_seq_len = key_states.shape[-2]
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
+
+        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class SiglipFlashAttention2(SiglipAttention):
+    """
+    Llama flash attention module. This module inherits from `LlamaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.is_causal = False  # Hack to make sure we don't use a causal mask
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        # cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        # if past_key_value is not None:
+        #     cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+        #     key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                "The input hidden states seems to be silently casted in float32, this might be related to the fact"
+                " you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights
+
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+
+        # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+        causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
+class SiglipEncoderLayer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.self_attn = (
+            SiglipAttention(config)
+            if not self._use_flash_attention_2
+            else SiglipFlashAttention2(config)
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(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,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class SiglipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SiglipVisionConfig
+    base_model_prefix = "siglip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+
+        if isinstance(module, SiglipVisionEmbeddings):
+            width = self.config.hidden_size
+            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
+        elif isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, SiglipAttention):
+            nn.init.normal_(module.q_proj.weight)
+            nn.init.normal_(module.k_proj.weight)
+            nn.init.normal_(module.v_proj.weight)
+            nn.init.normal_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, SiglipMLP):
+            nn.init.normal_(module.fc1.weight)
+            nn.init.normal_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+SIGLIP_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 ([`SiglipVisionConfig`]): 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.
+"""
+
+
+SIGLIP_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.
+"""
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
+class SiglipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`SiglipEncoderLayer`].
+    Args:
+        config: SiglipConfig
+    """
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    # Ignore copy
+    def forward(
+        self,
+        inputs_embeds,
+        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)
+            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 encoder_layer in 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,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        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
+        )
+
+@add_start_docstrings(
+    """The vision model from SigLIP without any head or projection on top.""",
+    SIGLIP_START_DOCSTRING
+)
+class SiglipVisionTransformer(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "pixel_values"
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SiglipVisionEmbeddings(config)
+        self.encoder = SiglipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipVisionConfig)
+    def forward(
+        self,
+        pixel_values,
+        patch_attention_mask: Optional[torch.BoolTensor] = None,
+        tgt_sizes: Optional[torch.IntTensor] = 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
+
+        batch_size = pixel_values.size(0)
+        if patch_attention_mask is None:
+            patch_attention_mask = torch.ones(
+                size=(
+                    batch_size,
+                    pixel_values.size(2) // self.config.patch_size,
+                    pixel_values.size(3) // self.config.patch_size,
+                ),
+                dtype=torch.bool,
+                device=pixel_values.device,
+            )
+
+        hidden_states = self.embeddings(pixel_values=pixel_values, patch_attention_mask=patch_attention_mask, tgt_sizes=tgt_sizes)
+
+        patch_attention_mask = patch_attention_mask.view(batch_size, -1)
+        # The call to `_upad_input` in `_flash_attention_forward` is expensive
+        # So when the `patch_attention_mask` is full of 1s (i.e. attending to the whole sequence),
+        # avoiding passing the attention_mask, which is equivalent to attending to the full sequence
+        if not torch.any(~patch_attention_mask):
+            attention_mask=None
+        else:
+            attention_mask = (
+                _prepare_4d_attention_mask(patch_attention_mask, hidden_states.dtype)
+                if not self._use_flash_attention_2
+                else patch_attention_mask
+            )
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=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.post_layernorm(last_hidden_state)
+
+        if not return_dict:
+            return (last_hidden_state, None) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=None,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

resampler.py

@@ -0,0 +1,782 @@
+from functools import partial
+from typing import Optional, Tuple
+import numpy as np
+import warnings
+
+import torch
+from torch import nn
+from torch import Tensor
+import torch.nn.functional as F
+from torch.nn.functional import *
+from torch.nn.modules.activation import *
+from torch.nn.init import trunc_normal_, constant_, xavier_normal_, xavier_uniform_
+
+from transformers.integrations import is_deepspeed_zero3_enabled
+
+def get_2d_sincos_pos_embed(embed_dim, image_size):
+    """
+    image_size: image_size or (image_height, image_width)
+    return:
+    pos_embed: [image_height, image_width, embed_dim]
+    """
+    if isinstance(image_size, int):
+        grid_h_size, grid_w_size = image_size, image_size
+    else:
+        grid_h_size, grid_w_size = image_size[0], image_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[0])  # (H, W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[1])  # (H, W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid_new(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (H, W)
+    out: (H, W, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    out = np.einsum('hw,d->hwd', pos, omega)  # (H, W, D/2), outer product
+
+    emb_sin = np.sin(out)  # (H, W, D/2)
+    emb_cos = np.cos(out)  # (H, W, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
+    return emb
+
+
+class Resampler(nn.Module):
+    """
+    A 2D perceiver-resampler network with one cross attention layers by
+       given learnable queries and 2d sincos pos_emb
+    Outputs:
+        A tensor with the shape of (batch_size, num_queries, embed_dim)
+    """
+
+    def __init__(
+            self,
+            num_queries,
+            embed_dim,
+            num_heads,
+            kv_dim=None,
+            norm_layer=partial(nn.LayerNorm, eps=1e-6),
+            adaptive=False,
+            max_size=(70, 70),
+    ):
+        super().__init__()
+        self.num_queries = num_queries
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.adaptive = adaptive
+        self.max_size = max_size
+
+        self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
+
+        if kv_dim is not None and kv_dim != embed_dim:
+            self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
+        else:
+            self.kv_proj = nn.Identity()
+
+        self.attn = MultiheadAttention(embed_dim, num_heads)
+        self.ln_q = norm_layer(embed_dim)
+        self.ln_kv = norm_layer(embed_dim)
+
+        self.ln_post = norm_layer(embed_dim)
+        self.proj = nn.Parameter((embed_dim ** -0.5) * torch.randn(embed_dim, embed_dim))
+
+        self._set_2d_pos_cache(self.max_size)
+
+    def _set_2d_pos_cache(self, max_size, device='cpu'):
+        if is_deepspeed_zero3_enabled():
+            device='cuda'
+        pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.embed_dim, max_size)).float().to(device)
+        self.register_buffer("pos_embed", pos_embed, persistent=False)
+
+    def _adjust_pos_cache(self, tgt_sizes, device):
+        max_h = torch.max(tgt_sizes[:, 0])
+        max_w = torch.max(tgt_sizes[:, 1])
+        if max_h > self.max_size[0] or max_w > self.max_size[1]:
+            self.max_size = [max(max_h, self.max_size[0]), max(max_w, self.max_size[1])]
+            self._set_2d_pos_cache(self.max_size, device)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(self, x, tgt_sizes=None):
+        assert x.shape[0] == tgt_sizes.shape[0]
+        bs = x.shape[0]
+
+        device = x.device
+        dtype = x.dtype
+
+        patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]
+
+        self._adjust_pos_cache(tgt_sizes, device=device)
+
+        max_patch_len = torch.max(patch_len)
+        key_padding_mask = torch.zeros((bs, max_patch_len), dtype=torch.bool, device=device)
+
+        pos_embed = []
+        for i in range(bs):
+            tgt_h, tgt_w = tgt_sizes[i]
+            pos_embed.append(self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype))  # patches * D
+            key_padding_mask[i, patch_len[i]:] = True
+
+        pos_embed = torch.nn.utils.rnn.pad_sequence(
+            pos_embed, batch_first=True, padding_value=0.0).permute(1, 0, 2)  # BLD => L * B * D
+
+        x = self.kv_proj(x)  # B * L * D
+        x = self.ln_kv(x).permute(1, 0, 2)  # L * B * D
+
+        q = self.ln_q(self.query)  # Q * D
+
+        out = self.attn(
+            self._repeat(q, bs),  # Q * B * D
+            x + pos_embed,  # L * B * D +  L * B * D
+            x,
+            key_padding_mask=key_padding_mask)[0]
+        #  out: Q * B * D
+        x = out.permute(1, 0, 2)  # B * Q * D
+
+        x = self.ln_post(x)
+        x = x @ self.proj
+        return x
+
+    def _repeat(self, query, N: int):
+        return query.unsqueeze(1).repeat(1, N, 1)
+
+
+class MultiheadAttention(nn.MultiheadAttention):
+    def __init__(self, embed_dim, num_heads, dropout=0., bias=True, add_bias_kv=False, 
+                 add_zero_attn=False, kdim=None, vdim=None, batch_first=False, device=None, dtype=None):
+        super().__init__(embed_dim, num_heads, dropout, bias, add_bias_kv, add_zero_attn, kdim, vdim, batch_first, device, dtype)
+
+        # rewrite out_proj layer，with nn.Linear
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(
+                self,
+                query: Tensor,
+                key: Tensor,
+                value: Tensor,
+                key_padding_mask: Optional[Tensor] = None,
+                need_weights: bool = True,
+                attn_mask: Optional[Tensor] = None,
+                average_attn_weights: bool = True,
+                is_causal : bool = False) -> Tuple[Tensor, Optional[Tensor]]:
+        why_not_fast_path = ''
+        if ((attn_mask is not None and torch.is_floating_point(attn_mask))
+           or (key_padding_mask is not None) and torch.is_floating_point(key_padding_mask)):
+            why_not_fast_path = "floating-point masks are not supported for fast path."
+
+        is_batched = query.dim() == 3
+
+        key_padding_mask = _canonical_mask(
+            mask=key_padding_mask,
+            mask_name="key_padding_mask",
+            other_type=F._none_or_dtype(attn_mask),
+            other_name="attn_mask",
+            target_type=query.dtype
+        )
+
+        attn_mask = _canonical_mask(
+            mask=attn_mask,
+            mask_name="attn_mask",
+            other_type=None,
+            other_name="",
+            target_type=query.dtype,
+            check_other=False,
+        )
+
+
+        if not is_batched:
+            why_not_fast_path = f"input not batched; expected query.dim() of 3 but got {query.dim()}"
+        elif query is not key or key is not value:
+            # When lifting this restriction, don't forget to either
+            # enforce that the dtypes all match or test cases where
+            # they don't!
+            why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
+        elif self.in_proj_bias is not None and query.dtype != self.in_proj_bias.dtype:
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
+        elif self.in_proj_weight is None:
+            why_not_fast_path = "in_proj_weight was None"
+        elif query.dtype != self.in_proj_weight.dtype:
+            # this case will fail anyway, but at least they'll get a useful error message.
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
+        elif self.training:
+            why_not_fast_path = "training is enabled"
+        elif (self.num_heads % 2) != 0:
+            why_not_fast_path = "self.num_heads is not even"
+        elif not self.batch_first:
+            why_not_fast_path = "batch_first was not True"
+        elif self.bias_k is not None:
+            why_not_fast_path = "self.bias_k was not None"
+        elif self.bias_v is not None:
+            why_not_fast_path = "self.bias_v was not None"
+        elif self.add_zero_attn:
+            why_not_fast_path = "add_zero_attn was enabled"
+        elif not self._qkv_same_embed_dim:
+            why_not_fast_path = "_qkv_same_embed_dim was not True"
+        elif query.is_nested and (key_padding_mask is not None or attn_mask is not None):
+            why_not_fast_path = "supplying both src_key_padding_mask and src_mask at the same time \
+                                 is not supported with NestedTensor input"
+        elif torch.is_autocast_enabled():
+            why_not_fast_path = "autocast is enabled"
+
+        if not why_not_fast_path:
+            tensor_args = (
+                query,
+                key,
+                value,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.out_proj.weight,
+                self.out_proj.bias,
+            )
+            # We have to use list comprehensions below because TorchScript does not support
+            # generator expressions.
+            if torch.overrides.has_torch_function(tensor_args):
+                why_not_fast_path = "some Tensor argument has_torch_function"
+            elif _is_make_fx_tracing():
+                why_not_fast_path = "we are running make_fx tracing"
+            elif not all(_check_arg_device(x) for x in tensor_args):
+                why_not_fast_path = ("some Tensor argument's device is neither one of "
+                                     f"cpu, cuda or {torch.utils.backend_registration._privateuse1_backend_name}")
+            elif torch.is_grad_enabled() and any(_arg_requires_grad(x) for x in tensor_args):
+                why_not_fast_path = ("grad is enabled and at least one of query or the "
+                                     "input/output projection weights or biases requires_grad")
+            if not why_not_fast_path:
+                merged_mask, mask_type = self.merge_masks(attn_mask, key_padding_mask, query)
+
+                if self.in_proj_bias is not None and self.in_proj_weight is not None:
+                    return torch._native_multi_head_attention(
+                        query,
+                        key,
+                        value,
+                        self.embed_dim,
+                        self.num_heads,
+                        self.in_proj_weight,
+                        self.in_proj_bias,
+                        self.out_proj.weight,
+                        self.out_proj.bias,
+                        merged_mask,
+                        need_weights,
+                        average_attn_weights,
+                        mask_type)
+
+        any_nested = query.is_nested or key.is_nested or value.is_nested
+        assert not any_nested, ("MultiheadAttention does not support NestedTensor outside of its fast path. " +
+                                f"The fast path was not hit because {why_not_fast_path}")
+
+        if self.batch_first and is_batched:
+            # make sure that the transpose op does not affect the "is" property
+            if key is value:
+                if query is key:
+                    query = key = value = query.transpose(1, 0)
+                else:
+                    query, key = (x.transpose(1, 0) for x in (query, key))
+                    value = key
+            else:
+                query, key, value = (x.transpose(1, 0) for x in (query, key, value))
+        
+        if not self._qkv_same_embed_dim:
+            attn_output, attn_output_weights = self.multi_head_attention_forward(
+                query, key, value, self.embed_dim, self.num_heads,
+                self.in_proj_weight, self.in_proj_bias,
+                self.bias_k, self.bias_v, self.add_zero_attn,
+                self.dropout, self.out_proj.weight, self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask, need_weights=need_weights,
+                attn_mask=attn_mask,
+                use_separate_proj_weight=True,
+                q_proj_weight=self.q_proj_weight, k_proj_weight=self.k_proj_weight,
+                v_proj_weight=self.v_proj_weight,
+                average_attn_weights=average_attn_weights,
+                is_causal=is_causal)
+        else:
+            attn_output, attn_output_weights = self.multi_head_attention_forward(
+                query, key, value, self.embed_dim, self.num_heads,
+                self.in_proj_weight, self.in_proj_bias,
+                self.bias_k, self.bias_v, self.add_zero_attn,
+                self.dropout, self.out_proj.weight, self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                average_attn_weights=average_attn_weights,
+                is_causal=is_causal)
+        if self.batch_first and is_batched:
+            return attn_output.transpose(1, 0), attn_output_weights
+        else:
+            return attn_output, attn_output_weights
+            
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Optional[Tensor],
+        in_proj_bias: Optional[Tensor],
+        bias_k: Optional[Tensor],
+        bias_v: Optional[Tensor],
+        add_zero_attn: bool,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Optional[Tensor],
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        use_separate_proj_weight: bool = False,
+        q_proj_weight: Optional[Tensor] = None,
+        k_proj_weight: Optional[Tensor] = None,
+        v_proj_weight: Optional[Tensor] = None,
+        static_k: Optional[Tensor] = None,
+        static_v: Optional[Tensor] = None,
+        average_attn_weights: bool = True,
+        is_causal: bool = False,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k, bias_v, out_proj_weight, out_proj_bias)
+    
+        is_batched = _mha_shape_check(query, key, value, key_padding_mask, attn_mask, num_heads)
+    
+        # For unbatched input, we unsqueeze at the expected batch-dim to pretend that the input
+        # is batched, run the computation and before returning squeeze the
+        # batch dimension so that the output doesn't carry this temporary batch dimension.
+        if not is_batched:
+            # unsqueeze if the input is unbatched
+            query = query.unsqueeze(1)
+            key = key.unsqueeze(1)
+            value = value.unsqueeze(1)
+            if key_padding_mask is not None:
+                key_padding_mask = key_padding_mask.unsqueeze(0)
+    
+        # set up shape vars
+        tgt_len, bsz, embed_dim = query.shape
+        src_len, _, _ = key.shape
+    
+        key_padding_mask = _canonical_mask(
+            mask=key_padding_mask,
+            mask_name="key_padding_mask",
+            other_type=_none_or_dtype(attn_mask),
+            other_name="attn_mask",
+            target_type=query.dtype
+        )
+    
+        if is_causal and attn_mask is None:
+            raise RuntimeError(
+                "Need attn_mask if specifying the is_causal hint. "
+                "You may use the Transformer module method "
+                "`generate_square_subsequent_mask` to create this mask."
+            )
+    
+        if is_causal and key_padding_mask is None and not need_weights:
+            # when we have a kpm or need weights, we need attn_mask
+            # Otherwise, we use the is_causal hint go as is_causal
+            # indicator to SDPA.
+            attn_mask = None
+        else:
+            attn_mask = _canonical_mask(
+                mask=attn_mask,
+                mask_name="attn_mask",
+                other_type=None,
+                other_name="",
+                target_type=query.dtype,
+                check_other=False,
+            )
+    
+            if key_padding_mask is not None:
+                # We have the attn_mask, and use that to merge kpm into it.
+                # Turn off use of is_causal hint, as the merged mask is no
+                # longer causal.
+                is_causal = False
+    
+        assert embed_dim == embed_dim_to_check, \
+            f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
+        if isinstance(embed_dim, torch.Tensor):
+            # embed_dim can be a tensor when JIT tracing
+            head_dim = embed_dim.div(num_heads, rounding_mode='trunc')
+        else:
+            head_dim = embed_dim // num_heads
+        assert head_dim * num_heads == embed_dim, f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
+        if use_separate_proj_weight:
+            # allow MHA to have different embedding dimensions when separate projection weights are used
+            assert key.shape[:2] == value.shape[:2], \
+                f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
+        else:
+            assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
+    
+        #
+        # compute in-projection
+        #
+        if not use_separate_proj_weight:
+            assert in_proj_weight is not None, "use_separate_proj_weight is False but in_proj_weight is None"
+            q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
+        else:
+            assert q_proj_weight is not None, "use_separate_proj_weight is True but q_proj_weight is None"
+            assert k_proj_weight is not None, "use_separate_proj_weight is True but k_proj_weight is None"
+            assert v_proj_weight is not None, "use_separate_proj_weight is True but v_proj_weight is None"
+            if in_proj_bias is None:
+                b_q = b_k = b_v = None
+            else:
+                b_q, b_k, b_v = in_proj_bias.chunk(3)
+            q, k, v = _in_projection(query, key, value, q_proj_weight, k_proj_weight, v_proj_weight, b_q, b_k, b_v)
+    
+        # prep attention mask
+    
+        if attn_mask is not None:
+            # ensure attn_mask's dim is 3
+            if attn_mask.dim() == 2:
+                correct_2d_size = (tgt_len, src_len)
+                if attn_mask.shape != correct_2d_size:
+                    raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
+                attn_mask = attn_mask.unsqueeze(0)
+            elif attn_mask.dim() == 3:
+                correct_3d_size = (bsz * num_heads, tgt_len, src_len)
+                if attn_mask.shape != correct_3d_size:
+                    raise RuntimeError(f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}.")
+            else:
+                raise RuntimeError(f"attn_mask's dimension {attn_mask.dim()} is not supported")
+    
+        # add bias along batch dimension (currently second)
+        if bias_k is not None and bias_v is not None:
+            assert static_k is None, "bias cannot be added to static key."
+            assert static_v is None, "bias cannot be added to static value."
+            k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
+            v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
+            if attn_mask is not None:
+                attn_mask = pad(attn_mask, (0, 1))
+            if key_padding_mask is not None:
+                key_padding_mask = pad(key_padding_mask, (0, 1))
+        else:
+            assert bias_k is None
+            assert bias_v is None
+    
+        #
+        # reshape q, k, v for multihead attention and make em batch first
+        #
+        q = q.view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
+        if static_k is None:
+            k = k.view(k.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+        else:
+            # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+            assert static_k.size(0) == bsz * num_heads, \
+                f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
+            assert static_k.size(2) == head_dim, \
+                f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
+            k = static_k
+        if static_v is None:
+            v = v.view(v.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+        else:
+            # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+            assert static_v.size(0) == bsz * num_heads, \
+                f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
+            assert static_v.size(2) == head_dim, \
+                f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
+            v = static_v
+    
+        # add zero attention along batch dimension (now first)
+        if add_zero_attn:
+            zero_attn_shape = (bsz * num_heads, 1, head_dim)
+            k = torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1)
+            v = torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1)
+            if attn_mask is not None:
+                attn_mask = pad(attn_mask, (0, 1))
+            if key_padding_mask is not None:
+                key_padding_mask = pad(key_padding_mask, (0, 1))
+    
+        # update source sequence length after adjustments
+        src_len = k.size(1)
+    
+        # merge key padding and attention masks
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (bsz, src_len), \
+                f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
+            key_padding_mask = key_padding_mask.view(bsz, 1, 1, src_len).   \
+                expand(-1, num_heads, -1, -1).reshape(bsz * num_heads, 1, src_len)
+            if attn_mask is None:
+                attn_mask = key_padding_mask
+            else:
+                attn_mask = attn_mask + key_padding_mask
+    
+        # adjust dropout probability
+        if not training:
+            dropout_p = 0.0
+    
+        #
+        # (deep breath) calculate attention and out projection
+        #
+    
+        if need_weights:
+            B, Nt, E = q.shape
+            q_scaled = q / math.sqrt(E)
+    
+            assert not (is_causal and attn_mask is None), "FIXME: is_causal not implemented for need_weights"
+    
+            if attn_mask is not None:
+                attn_output_weights = torch.baddbmm(attn_mask, q_scaled, k.transpose(-2, -1))
+            else:
+                attn_output_weights = torch.bmm(q_scaled, k.transpose(-2, -1))
+            attn_output_weights = softmax(attn_output_weights, dim=-1)
+            if dropout_p > 0.0:
+                attn_output_weights = dropout(attn_output_weights, p=dropout_p)
+    
+            attn_output = torch.bmm(attn_output_weights, v)
+    
+            attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
+            attn_output = self.out_proj(attn_output)
+            attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+    
+            # optionally average attention weights over heads
+            attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
+            if average_attn_weights:
+                attn_output_weights = attn_output_weights.mean(dim=1)
+    
+            if not is_batched:
+                # squeeze the output if input was unbatched
+                attn_output = attn_output.squeeze(1)
+                attn_output_weights = attn_output_weights.squeeze(0)
+            return attn_output, attn_output_weights
+        else:
+            # attn_mask can be either (L,S) or (N*num_heads, L, S)
+            # if attn_mask's shape is (1, L, S) we need to unsqueeze to (1, 1, L, S)
+            # in order to match the input for SDPA of (N, num_heads, L, S)
+            if attn_mask is not None:
+                if attn_mask.size(0) == 1 and attn_mask.dim() == 3:
+                    attn_mask = attn_mask.unsqueeze(0)
+                else:
+                    attn_mask = attn_mask.view(bsz, num_heads, -1, src_len)
+    
+            q = q.view(bsz, num_heads, tgt_len, head_dim)
+            k = k.view(bsz, num_heads, src_len, head_dim)
+            v = v.view(bsz, num_heads, src_len, head_dim)
+    
+            attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
+            attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
+    
+            attn_output = self.out_proj(attn_output)
+            attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+            if not is_batched:
+                # squeeze the output if input was unbatched
+                attn_output = attn_output.squeeze(1)
+            return attn_output, None
+
+
+def _mha_shape_check(query: Tensor, key: Tensor, value: Tensor,
+                     key_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor], num_heads: int):
+    # Verifies the expected shape for `query, `key`, `value`, `key_padding_mask` and `attn_mask`
+    # and returns if the input is batched or not.
+    # Raises an error if `query` is not 2-D (unbatched) or 3-D (batched) tensor.
+
+    # Shape check.
+    if query.dim() == 3:
+        # Batched Inputs
+        is_batched = True
+        assert key.dim() == 3 and value.dim() == 3, \
+            ("For batched (3-D) `query`, expected `key` and `value` to be 3-D"
+             f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
+        if key_padding_mask is not None:
+            assert key_padding_mask.dim() == 2, \
+                ("For batched (3-D) `query`, expected `key_padding_mask` to be `None` or 2-D"
+                 f" but found {key_padding_mask.dim()}-D tensor instead")
+        if attn_mask is not None:
+            assert attn_mask.dim() in (2, 3), \
+                ("For batched (3-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
+                 f" but found {attn_mask.dim()}-D tensor instead")
+    elif query.dim() == 2:
+        # Unbatched Inputs
+        is_batched = False
+        assert key.dim() == 2 and value.dim() == 2, \
+            ("For unbatched (2-D) `query`, expected `key` and `value` to be 2-D"
+             f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.dim() == 1, \
+                ("For unbatched (2-D) `query`, expected `key_padding_mask` to be `None` or 1-D"
+                 f" but found {key_padding_mask.dim()}-D tensor instead")
+
+        if attn_mask is not None:
+            assert attn_mask.dim() in (2, 3), \
+                ("For unbatched (2-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
+                 f" but found {attn_mask.dim()}-D tensor instead")
+            if attn_mask.dim() == 3:
+                expected_shape = (num_heads, query.shape[0], key.shape[0])
+                assert attn_mask.shape == expected_shape, \
+                    (f"Expected `attn_mask` shape to be {expected_shape} but got {attn_mask.shape}")
+    else:
+        raise AssertionError(
+            f"query should be unbatched 2D or batched 3D tensor but received {query.dim()}-D query tensor")
+
+    return is_batched
+
+
+def _canonical_mask(
+        mask: Optional[Tensor],
+        mask_name: str,
+        other_type: Optional[DType],
+        other_name: str,
+        target_type: DType,
+        check_other: bool = True,
+) -> Optional[Tensor]:
+
+    if mask is not None:
+        _mask_dtype = mask.dtype
+        _mask_is_float = torch.is_floating_point(mask)
+        if _mask_dtype != torch.bool and not _mask_is_float:
+            raise AssertionError(
+                f"only bool and floating types of {mask_name} are supported")
+        if check_other and other_type is not None:
+            if _mask_dtype != other_type:
+                warnings.warn(
+                    f"Support for mismatched {mask_name} and {other_name} "
+                    "is deprecated. Use same type for both instead."
+                )
+        if not _mask_is_float:
+            mask = (
+                torch.zeros_like(mask, dtype=target_type)
+                .masked_fill_(mask, float("-inf"))
+            )
+    return mask
+
+
+def _none_or_dtype(input: Optional[Tensor]) -> Optional[DType]:
+    if input is None:
+        return None
+    elif isinstance(input, torch.Tensor):
+        return input.dtype
+    raise RuntimeError("input to _none_or_dtype() must be None or torch.Tensor")
+
+def _in_projection_packed(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    w: Tensor,
+    b: Optional[Tensor] = None,
+) -> List[Tensor]:
+    r"""
+    Performs the in-projection step of the attention operation, using packed weights.
+    Output is a triple containing projection tensors for query, key and value.
+    Args:
+        q, k, v: query, key and value tensors to be projected. For self-attention,
+            these are typically the same tensor; for encoder-decoder attention,
+            k and v are typically the same tensor. (We take advantage of these
+            identities for performance if they are present.) Regardless, q, k and v
+            must share a common embedding dimension; otherwise their shapes may vary.
+        w: projection weights for q, k and v, packed into a single tensor. Weights
+            are packed along dimension 0, in q, k, v order.
+        b: optional projection biases for q, k and v, packed into a single tensor
+            in q, k, v order.
+    Shape:
+        Inputs:
+        - q: :math:`(..., E)` where E is the embedding dimension
+        - k: :math:`(..., E)` where E is the embedding dimension
+        - v: :math:`(..., E)` where E is the embedding dimension
+        - w: :math:`(E * 3, E)` where E is the embedding dimension
+        - b: :math:`E * 3` where E is the embedding dimension
+        Output:
+        - in output list :math:`[q', k', v']`, each output tensor will have the
+            same shape as the corresponding input tensor.
+    """
+    E = q.size(-1)
+    if k is v:
+        if q is k:
+            # self-attention
+            proj = linear(q, w, b)
+            # reshape to 3, E and not E, 3 is deliberate for better memory coalescing and keeping same order as chunk()
+            proj = proj.unflatten(-1, (3, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
+            return proj[0], proj[1], proj[2]
+        else:
+            # encoder-decoder attention
+            w_q, w_kv = w.split([E, E * 2])
+            if b is None:
+                b_q = b_kv = None
+            else:
+                b_q, b_kv = b.split([E, E * 2])
+            q_proj = linear(q, w_q, b_q)
+            kv_proj = linear(k, w_kv, b_kv)
+            # reshape to 2, E and not E, 2 is deliberate for better memory coalescing and keeping same order as chunk()
+            kv_proj = kv_proj.unflatten(-1, (2, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
+            return (q_proj, kv_proj[0], kv_proj[1])
+    else:
+        w_q, w_k, w_v = w.chunk(3)
+        if b is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = b.chunk(3)
+        return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
+
+
+def _in_projection(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    w_q: Tensor,
+    w_k: Tensor,
+    w_v: Tensor,
+    b_q: Optional[Tensor] = None,
+    b_k: Optional[Tensor] = None,
+    b_v: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor, Tensor]:
+    r"""
+    Performs the in-projection step of the attention operation. This is simply
+    a triple of linear projections, with shape constraints on the weights which
+    ensure embedding dimension uniformity in the projected outputs.
+    Output is a triple containing projection tensors for query, key and value.
+    Args:
+        q, k, v: query, key and value tensors to be projected.
+        w_q, w_k, w_v: weights for q, k and v, respectively.
+        b_q, b_k, b_v: optional biases for q, k and v, respectively.
+    Shape:
+        Inputs:
+        - q: :math:`(Qdims..., Eq)` where Eq is the query embedding dimension and Qdims are any
+            number of leading dimensions.
+        - k: :math:`(Kdims..., Ek)` where Ek is the key embedding dimension and Kdims are any
+            number of leading dimensions.
+        - v: :math:`(Vdims..., Ev)` where Ev is the value embedding dimension and Vdims are any
+            number of leading dimensions.
+        - w_q: :math:`(Eq, Eq)`
+        - w_k: :math:`(Eq, Ek)`
+        - w_v: :math:`(Eq, Ev)`
+        - b_q: :math:`(Eq)`
+        - b_k: :math:`(Eq)`
+        - b_v: :math:`(Eq)`
+        Output: in output triple :math:`(q', k', v')`,
+         - q': :math:`[Qdims..., Eq]`
+         - k': :math:`[Kdims..., Eq]`
+         - v': :math:`[Vdims..., Eq]`
+    """
+    Eq, Ek, Ev = q.size(-1), k.size(-1), v.size(-1)
+    assert w_q.shape == (Eq, Eq), f"expecting query weights shape of {(Eq, Eq)}, but got {w_q.shape}"
+    assert w_k.shape == (Eq, Ek), f"expecting key weights shape of {(Eq, Ek)}, but got {w_k.shape}"
+    assert w_v.shape == (Eq, Ev), f"expecting value weights shape of {(Eq, Ev)}, but got {w_v.shape}"
+    assert b_q is None or b_q.shape == (Eq,), f"expecting query bias shape of {(Eq,)}, but got {b_q.shape}"
+    assert b_k is None or b_k.shape == (Eq,), f"expecting key bias shape of {(Eq,)}, but got {b_k.shape}"
+    assert b_v is None or b_v.shape == (Eq,), f"expecting value bias shape of {(Eq,)}, but got {b_v.shape}"
+    return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v) 

rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e1fc6757d39c353f8c1b31430edf809bcb78e8fa70f565063f7ab91bde778ec
+size 14244

special_tokens_map.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3daa6a79e31deebf73b8ae5859068a107a4b0c4e745bb3bb99f6a562bbc194c
+size 1041

tokenization_minicpmv_fast.py

@@ -0,0 +1,66 @@
+from transformers.models.qwen2 import Qwen2TokenizerFast
+
+
+class MiniCPMVTokenizerFast(Qwen2TokenizerFast):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.im_start = "<image>"
+        self.im_end = "</image>"
+        self.ref_start = "<ref>"
+        self.ref_end = "</ref>"
+        self.box_start = "<box>"
+        self.box_end = "</box>"
+        self.quad_start = "<quad>"
+        self.quad_end = "</quad>"
+        self.slice_start = "<slice>"
+        self.slice_end = "</slice>"
+        self.im_id_start = "<image_id>"
+        self.im_id_end = "</image_id>"
+
+    @property
+    def eos_id(self):
+        return self.eos_token_id
+
+    @property
+    def bos_id(self):
+        return self.bos_token_id
+
+    @property
+    def unk_id(self):
+        return self.unk_token_id
+
+    @property
+    def im_start_id(self):
+        return self.convert_tokens_to_ids(self.im_start)
+
+    @property
+    def im_end_id(self):
+        return self.convert_tokens_to_ids(self.im_end)
+
+    @property
+    def slice_start_id(self):
+        return self.convert_tokens_to_ids(self.slice_start)
+    
+    @property
+    def slice_end_id(self):
+        return self.convert_tokens_to_ids(self.slice_end)
+
+    @property
+    def im_id_start_id(self):
+        return self.convert_tokens_to_ids(self.im_id_start)
+    
+    @property
+    def im_id_end_id(self):
+        return self.convert_tokens_to_ids(self.im_id_end)
+    
+    @property
+    def newline_id(self):
+        return self.convert_tokens_to_ids('\n')
+
+    @staticmethod
+    def escape(text: str) -> str:
+        return text
+
+    @staticmethod
+    def unescape(text: str) -> str:
+        return text

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:120c8f9b7a4f6fb721e3ffab5ab7863babd7f0ec706799192415dcd6a6e9fbbf
+size 7032034

tokenizer_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4d3bf8ad772cc0b30e75525c16f6952bf6d2ca4c3dff2b550b1b525420533ee
+size 5655

trainer_state.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c0d679de1f2001fd2d782f515a84731038b53c763bdbcf2cf251e8529bd8dca
+size 15816969

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0c91854272731588dfa1f756a02ccda3bfb0356692a9a7cd4f209bb4e14ec8b
+size 6712

vocab.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
+size 2776833

zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f1bbea82d83579e0f843af47db1458b0a53c126fde86cb542675ee08a34661f
+size 12844316057

zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)