Add files using upload-large-folder tool

LTA_openwebtext_dualt/logs/owt_dirichlet_len1024_Cv_to_2v_gumbel_sde_watch_fast/infer_0004000_gpu2.log

@@ -0,0 +1,12 @@
+[load] runs/lta_owt_dirichlet_len1024_Cv_to_2v_gbs512_4gpu_abspos_specialloss16_save1k_gumbelwatch_20260525_v2/step_0004000.pt
+[ckpt] step=4000
+[sde] generated 8/128
+[sde] generated 16/128
+[sde] generated 24/128
+[sde] generated 32/128
+[sde] generated 40/128
+[sde] generated 48/128
+[sde] generated 56/128
+[sde] generated 64/128
+[sde] generated 72/128
+[sde] generated 80/128

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deformable_detr/__init__.py

@@ -0,0 +1,31 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_deformable_detr import *
+    from .image_processing_deformable_detr import *
+    from .image_processing_pil_deformable_detr import *
+    from .modeling_deformable_detr import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deformable_detr/configuration_deformable_detr.py

@@ -0,0 +1,148 @@
+# Copyright 2022 SenseTime and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Deformable DETR model configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...backbone_utils import consolidate_backbone_kwargs_to_config
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+from ..auto import AutoConfig
+
+
+@auto_docstring(checkpoint="SenseTime/deformable-detr")
+@strict
+class DeformableDetrConfig(PreTrainedConfig):
+    r"""
+    num_queries (`int`, *optional*, defaults to 300):
+        Number of object queries, i.e. detection slots. This is the maximal number of objects
+        [`DeformableDetrModel`] can detect in a single image. In case `two_stage` is set to `True`, we use
+        `two_stage_num_proposals` instead.
+    return_intermediate (`bool`, *optional*, defaults to True):
+        Whether to return the intermediate state or not
+    position_embedding_type (`str`, *optional*, defaults to `"sine"`):
+        Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.
+    dilation (`bool`, *optional*, defaults to `False`):
+        Whether to replace stride with dilation in the last convolutional block (DC5). Only supported when
+        `use_timm_backbone` = `True`.
+    num_feature_levels (`int`, *optional*, defaults to 4):
+        The number of input feature levels.
+    encoder_n_points (`int`, *optional*, defaults to 4):
+        The number of sampled keys in each feature level for each attention head in the encoder.
+    decoder_n_points (`int`, *optional*, defaults to 4):
+        The number of sampled keys in each feature level for each attention head in the decoder.
+    two_stage (`bool`, *optional*, defaults to `False`):
+        Whether to apply a two-stage deformable DETR, where the region proposals are also generated by a variant of
+        Deformable DETR, which are further fed into the decoder for iterative bounding box refinement.
+    two_stage_num_proposals (`int`, *optional*, defaults to 300):
+        The number of region proposals to be generated, in case `two_stage` is set to `True`.
+    with_box_refine (`bool`, *optional*, defaults to `False`):
+        Whether to apply iterative bounding box refinement, where each decoder layer refines the bounding boxes
+        based on the predictions from the previous layer.
+    disable_custom_kernels (`bool`, *optional*, defaults to `False`):
+        Disable the use of custom CUDA and CPU kernels. This option is necessary for the ONNX export, as custom
+        kernels are not supported by PyTorch ONNX export.
+
+    Examples:
+
+    ```python
+    >>> from transformers import DeformableDetrConfig, DeformableDetrModel
+
+    >>> # Initializing a Deformable DETR SenseTime/deformable-detr style configuration
+    >>> configuration = DeformableDetrConfig()
+
+    >>> # Initializing a model (with random weights) from the SenseTime/deformable-detr style configuration
+    >>> model = DeformableDetrModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deformable_detr"
+    sub_configs = {"backbone_config": AutoConfig}
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    backbone_config: dict | PreTrainedConfig | None = None
+    num_channels: int = 3
+    num_queries: int = 300
+    max_position_embeddings: int = 1024
+    encoder_layers: int = 6
+    encoder_ffn_dim: int = 1024
+    encoder_attention_heads: int = 8
+    decoder_layers: int = 6
+    decoder_ffn_dim: int = 1024
+    decoder_attention_heads: int = 8
+    encoder_layerdrop: float | int = 0.0
+    is_encoder_decoder: bool = True
+    activation_function: str = "relu"
+    d_model: int = 256
+    dropout: float | int = 0.1
+    attention_dropout: float | int = 0.0
+    activation_dropout: float | int = 0.0
+    init_std: float = 0.02
+    init_xavier_std: float = 1.0
+    return_intermediate: bool = True
+    auxiliary_loss: bool = False
+    position_embedding_type: str = "sine"
+    dilation: bool = False
+    num_feature_levels: int = 4
+    encoder_n_points: int = 4
+    decoder_n_points: int = 4
+    two_stage: bool = False
+    two_stage_num_proposals: int = 300
+    with_box_refine: bool = False
+    class_cost: int = 1
+    bbox_cost: int = 5
+    giou_cost: int = 2
+    mask_loss_coefficient: int = 1
+    dice_loss_coefficient: int = 1
+    bbox_loss_coefficient: int = 5
+    giou_loss_coefficient: int = 2
+    eos_coefficient: float = 0.1
+    focal_alpha: float = 0.25
+    disable_custom_kernels: bool = False
+    tie_word_embeddings: bool = True
+
+    def __post_init__(self, **kwargs):
+        # Init timm backbone with hardcoded values for BC
+        timm_default_kwargs = {
+            "num_channels": 3,
+            "features_only": True,
+            "use_pretrained_backbone": False,
+            "out_indices": [2, 3, 4] if self.num_feature_levels > 1 else [4],
+        }
+        if self.dilation:
+            timm_default_kwargs["output_stride"] = 16
+
+        self.backbone_config, kwargs = consolidate_backbone_kwargs_to_config(
+            backbone_config=self.backbone_config,
+            default_backbone="resnet50",
+            default_config_type="resnet50",
+            default_config_kwargs={"out_features": ["stage4"]},
+            timm_default_kwargs=timm_default_kwargs,
+            **kwargs,
+        )
+
+        super().__post_init__(**kwargs)
+
+    def validate_architecture(self):
+        """Part of `@strict`-powered validation. Validates the architecture of the config."""
+        if self.two_stage is True and self.with_box_refine is False:
+            raise ValueError("If two_stage is True, with_box_refine must be True.")
+
+
+__all__ = ["DeformableDetrConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deformable_detr/image_processing_deformable_detr.py

@@ -0,0 +1,709 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deformable_detr/modular_deformable_detr.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_deformable_detr.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2022 SenseTime and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pathlib
+from typing import Any, Optional
+
+import torch
+from torchvision.io import read_image
+from torchvision.transforms.v2 import functional as tvF
+
+from ...image_processing_backends import TorchvisionBackend
+from ...image_processing_utils import BatchFeature, get_size_dict
+from ...image_transforms import (
+    center_to_corners_format,
+    corners_to_center_format,
+    get_size_with_aspect_ratio,
+    safe_squeeze,
+)
+from ...image_utils import (
+    IMAGENET_DEFAULT_MEAN,
+    IMAGENET_DEFAULT_STD,
+    AnnotationFormat,
+    AnnotationType,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+    get_image_size_for_max_height_width,
+    get_max_height_width,
+    validate_annotations,
+)
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring
+
+
+class DeformableDetrImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
+        Data format of the annotations. One of "coco_detection" or "coco_panoptic".
+    do_convert_annotations (`bool`, *optional*, defaults to `True`):
+        Controls whether to convert the annotations to the format expected by the DEFORMABLE_DETR model. Converts the
+        bounding boxes to the format `(center_x, center_y, width, height)` and in the range `[0, 1]`.
+        Can be overridden by the `do_convert_annotations` parameter in the `preprocess` method.
+    """
+
+    format: str | AnnotationFormat
+    do_convert_annotations: bool
+
+
+SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
+
+
+# inspired by https://github.com/facebookresearch/deformable_detr/blob/master/datasets/coco.py#L33
+def convert_coco_poly_to_mask(segmentations, height: int, width: int, device: torch.device) -> torch.Tensor:
+    """
+    Convert a COCO polygon annotation to a mask.
+
+    Args:
+        segmentations (`list[list[float]]`):
+            List of polygons, each polygon represented by a list of x-y coordinates.
+        height (`int`):
+            Height of the mask.
+        width (`int`):
+            Width of the mask.
+    """
+    try:
+        from pycocotools import mask as coco_mask
+    except ImportError:
+        raise ImportError("Pycocotools is not installed in your environment.")
+
+    masks = []
+    for polygons in segmentations:
+        rles = coco_mask.frPyObjects(polygons, height, width)
+        mask = coco_mask.decode(rles)
+        if len(mask.shape) < 3:
+            mask = mask[..., None]
+        mask = torch.as_tensor(mask, dtype=torch.uint8, device=device)
+        mask = torch.any(mask, axis=2)
+        masks.append(mask)
+    if masks:
+        masks = torch.stack(masks, axis=0)
+    else:
+        masks = torch.zeros((0, height, width), dtype=torch.uint8, device=device)
+
+    return masks
+
+
+# inspired by https://github.com/facebookresearch/deformable_detr/blob/master/datasets/coco.py#L50
+def prepare_coco_detection_annotation(
+    image,
+    target,
+    return_segmentation_masks: bool = False,
+    input_data_format: ChannelDimension | str | None = None,
+):
+    """
+    Convert the target in COCO format into the format expected by DEFORMABLE_DETR.
+    """
+    image_height, image_width = image.size()[-2:]
+
+    image_id = target["image_id"]
+    image_id = torch.as_tensor([image_id], dtype=torch.int64, device=image.device)
+
+    # Get all COCO annotations for the given image.
+    annotations = target["annotations"]
+    classes = []
+    area = []
+    boxes = []
+    keypoints = []
+    for obj in annotations:
+        if "iscrowd" not in obj or obj["iscrowd"] == 0:
+            classes.append(obj["category_id"])
+            area.append(obj["area"])
+            boxes.append(obj["bbox"])
+            if "keypoints" in obj:
+                keypoints.append(obj["keypoints"])
+
+    classes = torch.as_tensor(classes, dtype=torch.int64, device=image.device)
+    area = torch.as_tensor(area, dtype=torch.float32, device=image.device)
+    iscrowd = torch.zeros_like(classes, dtype=torch.int64, device=image.device)
+    # guard against no boxes via resizing
+    boxes = torch.as_tensor(boxes, dtype=torch.float32, device=image.device).reshape(-1, 4)
+    boxes[:, 2:] += boxes[:, :2]
+    boxes[:, 0::2] = boxes[:, 0::2].clip(min=0, max=image_width)
+    boxes[:, 1::2] = boxes[:, 1::2].clip(min=0, max=image_height)
+
+    keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
+
+    new_target = {
+        "image_id": image_id,
+        "class_labels": classes[keep],
+        "boxes": boxes[keep],
+        "area": area[keep],
+        "iscrowd": iscrowd[keep],
+        "orig_size": torch.as_tensor([int(image_height), int(image_width)], dtype=torch.int64, device=image.device),
+    }
+
+    if keypoints:
+        keypoints = torch.as_tensor(keypoints, dtype=torch.float32, device=image.device)
+        # Apply the keep mask here to filter the relevant annotations
+        keypoints = keypoints[keep]
+        num_keypoints = keypoints.shape[0]
+        keypoints = keypoints.reshape((-1, 3)) if num_keypoints else keypoints
+        new_target["keypoints"] = keypoints
+
+    if return_segmentation_masks:
+        segmentation_masks = [obj["segmentation"] for obj in annotations]
+        masks = convert_coco_poly_to_mask(segmentation_masks, image_height, image_width, device=image.device)
+        new_target["masks"] = masks[keep]
+
+    return new_target
+
+
+def masks_to_boxes(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Compute the bounding boxes around the provided panoptic segmentation masks.
+
+    Args:
+        masks: masks in format `[number_masks, height, width]` where N is the number of masks
+
+    Returns:
+        boxes: bounding boxes in format `[number_masks, 4]` in xyxy format
+    """
+    if masks.numel() == 0:
+        return torch.zeros((0, 4), device=masks.device)
+
+    h, w = masks.shape[-2:]
+    y = torch.arange(0, h, dtype=torch.float32, device=masks.device)
+    x = torch.arange(0, w, dtype=torch.float32, device=masks.device)
+    # see https://github.com/pytorch/pytorch/issues/50276
+    y, x = torch.meshgrid(y, x, indexing="ij")
+
+    x_mask = masks * torch.unsqueeze(x, 0)
+    x_max = x_mask.view(x_mask.shape[0], -1).max(-1)[0]
+    x_min = (
+        torch.where(masks, x.unsqueeze(0), torch.tensor(1e8, device=masks.device)).view(masks.shape[0], -1).min(-1)[0]
+    )
+
+    y_mask = masks * torch.unsqueeze(y, 0)
+    y_max = y_mask.view(y_mask.shape[0], -1).max(-1)[0]
+    y_min = (
+        torch.where(masks, y.unsqueeze(0), torch.tensor(1e8, device=masks.device)).view(masks.shape[0], -1).min(-1)[0]
+    )
+
+    return torch.stack([x_min, y_min, x_max, y_max], 1)
+
+
+# 2 functions below adapted from https://github.com/cocodataset/panopticapi/blob/master/panopticapi/utils.py
+# Copyright (c) 2018, Alexander Kirillov
+# All rights reserved.
+def rgb_to_id(color):
+    """
+    Converts RGB color to unique ID.
+    """
+    if isinstance(color, torch.Tensor) and len(color.shape) == 3:
+        if color.dtype == torch.uint8:
+            color = color.to(torch.int32)
+        return color[:, :, 0] + 256 * color[:, :, 1] + 256 * 256 * color[:, :, 2]
+    return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
+
+
+def prepare_coco_panoptic_annotation(
+    image: torch.Tensor,
+    target: dict,
+    masks_path: str | pathlib.Path,
+    return_masks: bool = True,
+    input_data_format: ChannelDimension | str = None,
+) -> dict:
+    """
+    Prepare a coco panoptic annotation for DEFORMABLE_DETR.
+    """
+    image_height, image_width = get_image_size(image, channel_dim=input_data_format)
+    annotation_path = pathlib.Path(masks_path) / target["file_name"]
+
+    new_target = {}
+    new_target["image_id"] = torch.as_tensor(
+        [target["image_id"] if "image_id" in target else target["id"]], dtype=torch.int64, device=image.device
+    )
+    new_target["size"] = torch.as_tensor([image_height, image_width], dtype=torch.int64, device=image.device)
+    new_target["orig_size"] = torch.as_tensor([image_height, image_width], dtype=torch.int64, device=image.device)
+
+    if "segments_info" in target:
+        masks = read_image(annotation_path).permute(1, 2, 0).to(dtype=torch.int32, device=image.device)
+        masks = rgb_to_id(masks)
+
+        ids = torch.as_tensor([segment_info["id"] for segment_info in target["segments_info"]], device=image.device)
+        masks = masks == ids[:, None, None]
+        masks = masks.to(torch.bool)
+        if return_masks:
+            new_target["masks"] = masks
+        new_target["boxes"] = masks_to_boxes(masks)
+        new_target["class_labels"] = torch.as_tensor(
+            [segment_info["category_id"] for segment_info in target["segments_info"]],
+            dtype=torch.int64,
+            device=image.device,
+        )
+        new_target["iscrowd"] = torch.as_tensor(
+            [segment_info["iscrowd"] for segment_info in target["segments_info"]],
+            dtype=torch.int64,
+            device=image.device,
+        )
+        new_target["area"] = torch.as_tensor(
+            [segment_info["area"] for segment_info in target["segments_info"]],
+            dtype=torch.float32,
+            device=image.device,
+        )
+
+    return new_target
+
+
+@auto_docstring
+class DeformableDetrImageProcessor(TorchvisionBackend):
+    valid_kwargs = DeformableDetrImageProcessorKwargs
+    resample = PILImageResampling.BILINEAR
+    image_mean = IMAGENET_DEFAULT_MEAN
+    image_std = IMAGENET_DEFAULT_STD
+    format = AnnotationFormat.COCO_DETECTION
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_pad = True
+    size = {"shortest_edge": 800, "longest_edge": 1333}
+    default_to_square = False
+    model_input_names = ["pixel_values", "pixel_mask"]
+
+    def __init__(self, **kwargs: Unpack[DeformableDetrImageProcessorKwargs]) -> None:
+        kwargs.setdefault("do_pad", kwargs.pop("pad_and_return_pixel_mask", self.do_pad))
+
+        size = kwargs.pop("size", None)
+        max_size = None if size is None else kwargs.pop("max_size", 1333)
+        size = size if size is not None else {"shortest_edge": 800, "longest_edge": 1333}
+        # Convert size dict for backwards compat with max_size parameter
+        kwargs["size"] = get_size_dict(size, max_size=max_size, default_to_square=False)
+
+        # Backwards compatibility
+        do_convert_annotations = kwargs.get("do_convert_annotations")
+        do_normalize = kwargs.get("do_normalize")
+        if do_convert_annotations is None and getattr(self, "do_convert_annotations", None) is None:
+            self.do_convert_annotations = do_normalize if do_normalize is not None else self.do_normalize
+
+        super().__init__(**kwargs)
+
+    def prepare_annotation(
+        self,
+        image: torch.Tensor,
+        target: dict,
+        format: AnnotationFormat | None = None,
+        return_segmentation_masks: bool | None = None,
+        masks_path: str | pathlib.Path | None = None,
+        input_data_format: str | ChannelDimension | None = None,
+    ) -> dict:
+        """
+        Prepare an annotation for feeding into DEFORMABLE_DETR model.
+        """
+        format = format if format is not None else self.format
+
+        if format == AnnotationFormat.COCO_DETECTION:
+            return_segmentation_masks = False if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_detection_annotation(
+                image, target, return_segmentation_masks, input_data_format=input_data_format
+            )
+        elif format == AnnotationFormat.COCO_PANOPTIC:
+            return_segmentation_masks = True if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_panoptic_annotation(
+                image,
+                target,
+                masks_path=masks_path,
+                return_masks=return_segmentation_masks,
+                input_data_format=input_data_format,
+            )
+        else:
+            raise ValueError(f"Format {format} is not supported.")
+        return target
+
+    def resize(
+        self,
+        image: torch.Tensor,
+        size: SizeDict,
+        resample: Optional["PILImageResampling | tvF.InterpolationMode | int"] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        Resize the image to the given size. Size can be `min_size` (scalar) or `(height, width)` tuple. If size is an
+        int, smaller edge of the image will be matched to this number.
+
+        Args:
+            image (`torch.Tensor`):
+                Image to resize.
+            size (`SizeDict`):
+                Size of the image's `(height, width)` dimensions after resizing. Available options are:
+                    - `{"height": int, "width": int}`: The image will be resized to the exact size `(height, width)`.
+                        Do NOT keep the aspect ratio.
+                    - `{"shortest_edge": int, "longest_edge": int}`: The image will be resized to a maximum size respecting
+                        the aspect ratio and keeping the shortest edge less or equal to `shortest_edge` and the longest edge
+                        less or equal to `longest_edge`.
+                    - `{"max_height": int, "max_width": int}`: The image will be resized to the maximum size respecting the
+                        aspect ratio and keeping the height less or equal to `max_height` and the width less or equal to
+                        `max_width`.
+            resample (`PILImageResampling | tvF.InterpolationMode | int`, *optional*, defaults to `PILImageResampling.BILINEAR`):
+                Resampling filter to use if resizing the image.
+        """
+        if size.shortest_edge and size.longest_edge:
+            # Resize the image so that the shortest edge or the longest edge is of the given size
+            # while maintaining the aspect ratio of the original image.
+            new_size = get_size_with_aspect_ratio(image.shape[-2:], size.shortest_edge, size.longest_edge)
+        elif size.max_height and size.max_width:
+            new_size = get_image_size_for_max_height_width(image.shape[-2:], size.max_height, size.max_width)
+        elif size.height and size.width:
+            new_size = (size.height, size.width)
+        else:
+            raise ValueError(
+                f"Size must contain 'height' and 'width' keys or 'shortest_edge' and 'longest_edge' keys. Got {size}."
+            )
+
+        image = super().resize(
+            image, size=SizeDict(height=new_size[0], width=new_size[1]), resample=resample, **kwargs
+        )
+        return image
+
+    def resize_annotation(
+        self,
+        annotation: dict[str, Any],
+        orig_size: tuple[int, int],
+        target_size: tuple[int, int],
+        threshold: float = 0.5,
+        resample: Optional["PILImageResampling | tvF.InterpolationMode | int"] = PILImageResampling.NEAREST,
+    ):
+        """
+        Resizes an annotation to a target size.
+
+        Args:
+            annotation (`dict[str, Any]`):
+                The annotation dictionary.
+            orig_size (`tuple[int, int]`):
+                The original size of the input image.
+            target_size (`tuple[int, int]`):
+                The target size of the image, as returned by the preprocessing `resize` step.
+            threshold (`float`, *optional*, defaults to 0.5):
+                The threshold used to binarize the segmentation masks.
+            resample (`PILImageResampling | tvF.InterpolationMode | int`, defaults to `tvF.InterpolationMode.NEAREST_EXACT`):
+                The resampling filter to use when resizing the masks.
+        """
+        ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
+
+        new_annotation = {}
+        new_annotation["size"] = target_size
+
+        for key, value in annotation.items():
+            if key == "boxes":
+                boxes = value
+                scaled_boxes = boxes * torch.as_tensor(
+                    [ratio_width, ratio_height, ratio_width, ratio_height], dtype=torch.float32, device=boxes.device
+                )
+                new_annotation["boxes"] = scaled_boxes
+            elif key == "area":
+                area = value
+                scaled_area = area * (ratio_width * ratio_height)
+                new_annotation["area"] = scaled_area
+            elif key == "masks":
+                masks = value[:, None]
+                masks = [
+                    super(DeformableDetrImageProcessor, self).resize(
+                        mask, size=SizeDict(height=target_size[0], width=target_size[1]), resample=resample
+                    )
+                    for mask in masks
+                ]
+                masks = torch.stack(masks).to(torch.float32)
+                masks = masks[:, 0] > threshold
+                new_annotation["masks"] = masks
+            elif key == "size":
+                new_annotation["size"] = target_size
+            else:
+                new_annotation[key] = value
+
+        return new_annotation
+
+    def normalize_annotation(self, annotation: dict, image_size: tuple[int, int]) -> dict:
+        image_height, image_width = image_size
+        norm_annotation = {}
+        for key, value in annotation.items():
+            if key == "boxes":
+                boxes = value
+                boxes = corners_to_center_format(boxes)
+                boxes /= torch.as_tensor(
+                    [image_width, image_height, image_width, image_height], dtype=torch.float32, device=boxes.device
+                )
+                norm_annotation[key] = boxes
+            else:
+                norm_annotation[key] = value
+        return norm_annotation
+
+    def _update_annotation_for_padded_image(
+        self,
+        annotation: dict,
+        input_image_size: tuple[int, int],
+        output_image_size: tuple[int, int],
+        padding,
+        update_bboxes,
+    ) -> dict:
+        """
+        Update the annotation for a padded image.
+        """
+        new_annotation = {}
+        new_annotation["size"] = output_image_size
+        ratio_height, ratio_width = (input / output for output, input in zip(output_image_size, input_image_size))
+
+        for key, value in annotation.items():
+            if key == "masks":
+                masks = value
+                masks = tvF.pad(
+                    masks,
+                    padding,
+                    fill=0,
+                )
+                masks = safe_squeeze(masks, 1)
+                new_annotation["masks"] = masks
+            elif key == "boxes" and update_bboxes:
+                boxes = value
+                boxes *= torch.as_tensor([ratio_width, ratio_height, ratio_width, ratio_height], device=boxes.device)
+                new_annotation["boxes"] = boxes
+            elif key == "size":
+                new_annotation["size"] = output_image_size
+            else:
+                new_annotation[key] = value
+        return new_annotation
+
+    def pad(
+        self,
+        image: torch.Tensor,
+        padded_size: tuple[int, int],
+        annotation: dict[str, Any] | None = None,
+        update_bboxes: bool = True,
+        fill: int = 0,
+    ):
+        original_size = image.size()[-2:]
+        padding_bottom = padded_size[0] - original_size[0]
+        padding_right = padded_size[1] - original_size[1]
+        if padding_bottom < 0 or padding_right < 0:
+            raise ValueError(
+                f"Padding dimensions are negative. Please make sure that the padded size is larger than the "
+                f"original size. Got padded size: {padded_size}, original size: {original_size}."
+            )
+        if original_size != padded_size:
+            padding = [0, 0, padding_right, padding_bottom]
+            image = tvF.pad(image, padding, fill=fill)
+            if annotation is not None:
+                annotation = self._update_annotation_for_padded_image(
+                    annotation, original_size, padded_size, padding, update_bboxes
+                )
+
+        # Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
+        pixel_mask = torch.zeros(padded_size, dtype=torch.int64, device=image.device)
+        pixel_mask[: original_size[0], : original_size[1]] = 1
+
+        return image, pixel_mask, annotation
+
+    @auto_docstring
+    def preprocess(
+        self,
+        images: ImageInput,
+        annotations: AnnotationType | list[AnnotationType] | None = None,
+        return_segmentation_masks: bool | None = None,
+        masks_path: str | pathlib.Path | None = None,
+        **kwargs: Unpack[DeformableDetrImageProcessorKwargs],
+    ) -> BatchFeature:
+        r"""
+        annotations (`AnnotationType` or `list[AnnotationType]`, *optional*):
+            Annotations to transform according to the padding that is applied to the images.
+        return_segmentation_masks (`bool`, *optional*, defaults to `self.return_segmentation_masks`):
+            Whether to return segmentation masks.
+        masks_path (`str` or `pathlib.Path`, *optional*):
+            Path to the directory containing the segmentation masks.
+        """
+        return super().preprocess(images, annotations, return_segmentation_masks, masks_path, **kwargs)
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        annotations: AnnotationType | list[AnnotationType] | None,
+        return_segmentation_masks: bool,
+        masks_path: str | pathlib.Path | None,
+        do_resize: bool,
+        size: SizeDict,
+        resample: "PILImageResampling | tvF.InterpolationMode | int | None",
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        do_convert_annotations: bool,
+        image_mean: float | list[float] | None,
+        image_std: float | list[float] | None,
+        do_pad: bool,
+        pad_size: SizeDict | None,
+        format: str | AnnotationFormat | None,
+        return_tensors: str | TensorType | None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Preprocess an image or a batch of images so that it can be used by the model.
+        """
+        if annotations is not None and isinstance(annotations, dict):
+            annotations = [annotations]
+
+        if annotations is not None and len(images) != len(annotations):
+            raise ValueError(
+                f"The number of images ({len(images)}) and annotations ({len(annotations)}) do not match."
+            )
+
+        format = AnnotationFormat(format)
+        if annotations is not None:
+            validate_annotations(format, SUPPORTED_ANNOTATION_FORMATS, annotations)
+
+        if (
+            masks_path is not None
+            and format == AnnotationFormat.COCO_PANOPTIC
+            and not isinstance(masks_path, (pathlib.Path, str))
+        ):
+            raise ValueError(
+                "The path to the directory containing the mask PNG files should be provided as a"
+                f" `pathlib.Path` or string object, but is {type(masks_path)} instead."
+            )
+
+        data = {}
+
+        processed_images = []
+        processed_annotations = []
+        pixel_masks = []  # Initialize pixel_masks here
+        for image, annotation in zip(images, annotations if annotations is not None else [None] * len(images)):
+            # prepare (COCO annotations as a list of Dict -> DEFORMABLE_DETR target as a single Dict per image)
+            if annotations is not None:
+                annotation = self.prepare_annotation(
+                    image,
+                    annotation,
+                    format,
+                    return_segmentation_masks=return_segmentation_masks,
+                    masks_path=masks_path,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+
+            if do_resize:
+                resized_image = self.resize(image, size=size, resample=resample)
+                if annotations is not None:
+                    annotation = self.resize_annotation(
+                        annotation,
+                        orig_size=image.size()[-2:],
+                        target_size=resized_image.size()[-2:],
+                    )
+                image = resized_image
+            # Fused rescale and normalize
+            image = self.rescale_and_normalize(image, do_rescale, rescale_factor, do_normalize, image_mean, image_std)
+            if do_convert_annotations and annotations is not None:
+                annotation = self.normalize_annotation(annotation, get_image_size(image, ChannelDimension.FIRST))
+
+            processed_images.append(image)
+            processed_annotations.append(annotation)
+        images = processed_images
+        annotations = processed_annotations if annotations is not None else None
+
+        if do_pad:
+            # depends on all resized image shapes so we need another loop
+            if pad_size is not None:
+                padded_size = (pad_size.height, pad_size.width)
+            else:
+                padded_size = get_max_height_width(images)
+
+            padded_images = []
+            padded_annotations = []
+            for image, annotation in zip(images, annotations if annotations is not None else [None] * len(images)):
+                # Pads images and returns their mask: {'pixel_values': ..., 'pixel_mask': ...}
+                if padded_size == image.size()[-2:]:
+                    padded_images.append(image)
+                    pixel_masks.append(torch.ones(padded_size, dtype=torch.int64, device=image.device))
+                    padded_annotations.append(annotation)
+                    continue
+                image, pixel_mask, annotation = self.pad(
+                    image, padded_size, annotation=annotation, update_bboxes=do_convert_annotations
+                )
+                padded_images.append(image)
+                padded_annotations.append(annotation)
+                pixel_masks.append(pixel_mask)
+            images = padded_images
+            annotations = padded_annotations if annotations is not None else None
+            data.update({"pixel_mask": torch.stack(pixel_masks, dim=0)})
+
+        data.update({"pixel_values": torch.stack(images, dim=0)})
+        encoded_inputs = BatchFeature(data, tensor_type=return_tensors)
+        if annotations is not None:
+            encoded_inputs["labels"] = [
+                BatchFeature(annotation, tensor_type=return_tensors) for annotation in annotations
+            ]
+        return encoded_inputs
+
+    def post_process_object_detection(
+        self, outputs, threshold: float = 0.5, target_sizes: TensorType | list[tuple] = None, top_k: int = 100
+    ):
+        """
+        Converts the raw output of [`DeformableDetrForObjectDetection`] into final bounding boxes in (top_left_x,
+        top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.
+
+        Args:
+            outputs ([`DetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*):
+                Score threshold to keep object detection predictions.
+            target_sizes (`torch.Tensor` or `list[tuple[int, int]]`, *optional*):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`tuple[int, int]`) containing the target size
+                (height, width) of each image in the batch. If left to None, predictions will not be resized.
+            top_k (`int`, *optional*, defaults to 100):
+                Keep only top k bounding boxes before filtering by thresholding.
+
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if target_sizes is not None:
+            if len(out_logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+        prob = out_logits.sigmoid()
+        prob = prob.view(out_logits.shape[0], -1)
+        k_value = min(top_k, prob.size(1))
+        topk_values, topk_indexes = torch.topk(prob, k_value, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            if isinstance(target_sizes, list):
+                img_h = torch.Tensor([i[0] for i in target_sizes])
+                img_w = torch.Tensor([i[1] for i in target_sizes])
+            else:
+                img_h, img_w = target_sizes.unbind(1)
+            scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1).to(boxes.device)
+            boxes = boxes * scale_fct[:, None, :]
+
+        results = []
+        for s, l, b in zip(scores, labels, boxes):
+            score = s[s > threshold]
+            label = l[s > threshold]
+            box = b[s > threshold]
+            results.append({"scores": score, "labels": label, "boxes": box})
+
+        return results
+
+
+__all__ = ["DeformableDetrImageProcessor"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deformable_detr/image_processing_pil_deformable_detr.py

@@ -0,0 +1,738 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deformable_detr/modular_deformable_detr.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_deformable_detr.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2022 SenseTime and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import pathlib
+from typing import Any, Optional
+
+import numpy as np
+
+from ...image_processing_backends import PilBackend
+from ...image_processing_utils import BatchFeature
+from ...image_transforms import (
+    PaddingMode,
+    center_to_corners_format,
+    corners_to_center_format,
+    get_size_with_aspect_ratio,
+    pad,
+    resize,
+    safe_squeeze,
+)
+from ...image_utils import (
+    IMAGENET_DEFAULT_MEAN,
+    IMAGENET_DEFAULT_STD,
+    AnnotationFormat,
+    AnnotationType,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+    get_image_size_for_max_height_width,
+    get_max_height_width,
+    validate_annotations,
+)
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring, is_torch_available, is_vision_available
+from ...utils.import_utils import requires, requires_backends
+
+
+if is_vision_available():
+    import PIL.Image
+if is_torch_available():
+    import torch
+
+SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
+
+
+class DeformableDetrImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
+        Data format of the annotations. One of "coco_detection" or "coco_panoptic".
+    do_convert_annotations (`bool`, *optional*, defaults to `True`):
+        Controls whether to convert the annotations to the format expected by the DEFORMABLE_DETR model. Converts the
+        bounding boxes to the format `(center_x, center_y, width, height)` and in the range `[0, 1]`.
+        Can be overridden by the `do_convert_annotations` parameter in the `preprocess` method.
+    """
+
+    format: str | AnnotationFormat
+    do_convert_annotations: bool
+
+
+# inspired by https://github.com/facebookresearch/deformable_detr/blob/master/datasets/coco.py#L33
+def convert_coco_poly_to_mask(segmentations, height: int, width: int) -> np.ndarray:
+    """
+    Convert a COCO polygon annotation to a mask.
+
+    Args:
+        segmentations (`list[list[float]]`):
+            List of polygons, each polygon represented by a list of x-y coordinates.
+        height (`int`):
+            Height of the mask.
+        width (`int`):
+            Width of the mask.
+    """
+    try:
+        from pycocotools import mask as coco_mask
+    except ImportError:
+        raise ImportError("Pycocotools is not installed in your environment.")
+
+    masks = []
+    for polygons in segmentations:
+        rles = coco_mask.frPyObjects(polygons, height, width)
+        mask = coco_mask.decode(rles)
+        if len(mask.shape) < 3:
+            mask = mask[..., None]
+        mask = np.asarray(mask, dtype=np.uint8)
+        mask = np.any(mask, axis=2)
+        masks.append(mask)
+    if masks:
+        masks = np.stack(masks, axis=0)
+    else:
+        masks = np.zeros((0, height, width), dtype=np.uint8)
+
+    return masks
+
+
+# inspired by https://github.com/facebookresearch/deformable_detr/blob/master/datasets/coco.py#L50
+def prepare_coco_detection_annotation(
+    image,
+    target,
+    return_segmentation_masks: bool = False,
+    input_data_format: ChannelDimension | str | None = None,
+):
+    """
+    Convert the target in COCO format into the format expected by DEFORMABLE_DETR.
+    """
+    image_height, image_width = get_image_size(image, channel_dim=input_data_format)
+
+    image_id = target["image_id"]
+    image_id = np.asarray([image_id], dtype=np.int64)
+
+    # Get all COCO annotations for the given image.
+    annotations = target["annotations"]
+    annotations = [obj for obj in annotations if "iscrowd" not in obj or obj["iscrowd"] == 0]
+
+    classes = [obj["category_id"] for obj in annotations]
+    classes = np.asarray(classes, dtype=np.int64)
+
+    # for conversion to coco api
+    area = np.asarray([obj["area"] for obj in annotations], dtype=np.float32)
+    iscrowd = np.asarray([obj.get("iscrowd", 0) for obj in annotations], dtype=np.int64)
+
+    boxes = [obj["bbox"] for obj in annotations]
+    # guard against no boxes via resizing
+    boxes = np.asarray(boxes, dtype=np.float32).reshape(-1, 4)
+    boxes[:, 2:] += boxes[:, :2]
+    boxes[:, 0::2] = boxes[:, 0::2].clip(min=0, max=image_width)
+    boxes[:, 1::2] = boxes[:, 1::2].clip(min=0, max=image_height)
+
+    keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
+
+    new_target = {}
+    new_target["image_id"] = image_id
+    new_target["class_labels"] = classes[keep]
+    new_target["boxes"] = boxes[keep]
+    new_target["area"] = area[keep]
+    new_target["iscrowd"] = iscrowd[keep]
+    new_target["orig_size"] = np.asarray([int(image_height), int(image_width)], dtype=np.int64)
+
+    if annotations and "keypoints" in annotations[0]:
+        keypoints = [obj["keypoints"] for obj in annotations]
+        # Converting the filtered keypoints list to a numpy array
+        keypoints = np.asarray(keypoints, dtype=np.float32)
+        # Apply the keep mask here to filter the relevant annotations
+        keypoints = keypoints[keep]
+        num_keypoints = keypoints.shape[0]
+        keypoints = keypoints.reshape((-1, 3)) if num_keypoints else keypoints
+        new_target["keypoints"] = keypoints
+
+    if return_segmentation_masks:
+        segmentation_masks = [obj["segmentation"] for obj in annotations]
+        masks = convert_coco_poly_to_mask(segmentation_masks, image_height, image_width)
+        new_target["masks"] = masks[keep]
+
+    return new_target
+
+
+def masks_to_boxes(masks: np.ndarray) -> np.ndarray:
+    """
+    Compute the bounding boxes around the provided panoptic segmentation masks.
+
+    Args:
+        masks: masks in format `[number_masks, height, width]` where N is the number of masks
+
+    Returns:
+        boxes: bounding boxes in format `[number_masks, 4]` in xyxy format
+    """
+    if masks.size == 0:
+        return np.zeros((0, 4))
+
+    h, w = masks.shape[-2:]
+    y = np.arange(0, h, dtype=np.float32)
+    x = np.arange(0, w, dtype=np.float32)
+    # see https://github.com/pytorch/pytorch/issues/50276
+    y, x = np.meshgrid(y, x, indexing="ij")
+
+    x_mask = masks * np.expand_dims(x, axis=0)
+    x_max = x_mask.reshape(x_mask.shape[0], -1).max(-1)
+    x = np.ma.array(x_mask, mask=~(np.array(masks, dtype=bool)))
+    x_min = x.filled(fill_value=1e8)
+    x_min = x_min.reshape(x_min.shape[0], -1).min(-1)
+
+    y_mask = masks * np.expand_dims(y, axis=0)
+    y_max = y_mask.reshape(x_mask.shape[0], -1).max(-1)
+    y = np.ma.array(y_mask, mask=~(np.array(masks, dtype=bool)))
+    y_min = y.filled(fill_value=1e8)
+    y_min = y_min.reshape(y_min.shape[0], -1).min(-1)
+
+    return np.stack([x_min, y_min, x_max, y_max], 1)
+
+
+# 2 functions below adapted from https://github.com/cocodataset/panopticapi/blob/master/panopticapi/utils.py
+# Copyright (c) 2018, Alexander Kirillov
+# All rights reserved.
+def rgb_to_id(color):
+    """
+    Converts RGB color to unique ID.
+    """
+    if isinstance(color, np.ndarray) and len(color.shape) == 3:
+        if color.dtype == np.uint8:
+            color = color.astype(np.int32)
+        return color[:, :, 0] + 256 * color[:, :, 1] + 256 * 256 * color[:, :, 2]
+    return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
+
+
+def prepare_coco_panoptic_annotation(
+    image: np.ndarray,
+    target: dict,
+    masks_path: str | pathlib.Path,
+    return_masks: bool = True,
+    input_data_format: ChannelDimension | str = None,
+) -> dict:
+    """
+    Prepare a coco panoptic annotation for DEFORMABLE_DETR.
+    """
+    image_height, image_width = get_image_size(image, channel_dim=input_data_format)
+    annotation_path = pathlib.Path(masks_path) / target["file_name"]
+
+    new_target = {}
+    new_target["image_id"] = np.asarray([target["image_id"] if "image_id" in target else target["id"]], dtype=np.int64)
+    new_target["size"] = np.asarray([image_height, image_width], dtype=np.int64)
+    new_target["orig_size"] = np.asarray([image_height, image_width], dtype=np.int64)
+
+    if "segments_info" in target:
+        masks = np.asarray(PIL.Image.open(annotation_path), dtype=np.uint32)
+        masks = rgb_to_id(masks)
+
+        ids = np.array([segment_info["id"] for segment_info in target["segments_info"]])
+        masks = masks == ids[:, None, None]
+        masks = masks.astype(np.uint8)
+        if return_masks:
+            new_target["masks"] = masks
+        new_target["boxes"] = masks_to_boxes(masks)
+        new_target["class_labels"] = np.array(
+            [segment_info["category_id"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["iscrowd"] = np.asarray(
+            [segment_info["iscrowd"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["area"] = np.asarray(
+            [segment_info["area"] for segment_info in target["segments_info"]], dtype=np.float32
+        )
+
+    return new_target
+
+
+@auto_docstring
+class DeformableDetrImageProcessorPil(PilBackend):
+    resample = PILImageResampling.BILINEAR
+    image_mean = IMAGENET_DEFAULT_MEAN
+    image_std = IMAGENET_DEFAULT_STD
+    format = AnnotationFormat.COCO_DETECTION
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_pad = True
+    size = {"shortest_edge": 800, "longest_edge": 1333}
+    default_to_square = False
+    model_input_names = ["pixel_values", "pixel_mask"]
+    valid_kwargs = DeformableDetrImageProcessorKwargs
+
+    def __init__(self, **kwargs: Unpack[DeformableDetrImageProcessorKwargs]) -> None:
+        kwargs.setdefault("do_pad", kwargs.pop("pad_and_return_pixel_mask", self.do_pad))
+
+        size = kwargs.pop("size", None)
+        max_size = None if size is None else kwargs.pop("max_size", 1333)
+        size = size if size is not None else {"shortest_edge": 800, "longest_edge": 1333}
+        # Convert size dict for backwards compat with max_size parameter
+        if size is not None:
+            from ...image_processing_utils import get_size_dict
+
+            kwargs["size"] = get_size_dict(size, max_size=max_size, default_to_square=False)
+
+        # Backwards compatibility
+        do_convert_annotations = kwargs.get("do_convert_annotations")
+        do_normalize = kwargs.get("do_normalize")
+        if do_convert_annotations is None and getattr(self, "do_convert_annotations", None) is None:
+            self.do_convert_annotations = do_normalize if do_normalize is not None else self.do_normalize
+
+        super().__init__(**kwargs)
+
+    def prepare_annotation(
+        self,
+        image: np.ndarray,
+        target: dict,
+        format: AnnotationFormat | None = None,
+        return_segmentation_masks: bool | None = None,
+        masks_path: str | pathlib.Path | None = None,
+        input_data_format: str | ChannelDimension | None = None,
+    ) -> dict:
+        """
+        Prepare an annotation for feeding into DEFORMABLE_DETR model.
+        """
+        format = format if format is not None else self.format
+
+        if format == AnnotationFormat.COCO_DETECTION:
+            return_segmentation_masks = False if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_detection_annotation(
+                image, target, return_segmentation_masks, input_data_format=input_data_format
+            )
+        elif format == AnnotationFormat.COCO_PANOPTIC:
+            return_segmentation_masks = True if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_panoptic_annotation(
+                image,
+                target,
+                masks_path=masks_path,
+                return_masks=return_segmentation_masks,
+                input_data_format=input_data_format,
+            )
+        else:
+            raise ValueError(f"Format {format} is not supported.")
+        return target
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: SizeDict,
+        resample: Optional["PILImageResampling"] = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Resize the image to the given size. Size can be `min_size` (scalar) or `(height, width)` tuple. If size is an
+        int, smaller edge of the image will be matched to this number.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`SizeDict`):
+                Size of the image's `(height, width)` dimensions after resizing. Available options are:
+                    - `{"height": int, "width": int}`: The image will be resized to the exact size `(height, width)`.
+                        Do NOT keep the aspect ratio.
+                    - `{"shortest_edge": int, "longest_edge": int}`: The image will be resized to a maximum size respecting
+                        the aspect ratio and keeping the shortest edge less or equal to `shortest_edge` and the longest edge
+                        less or equal to `longest_edge`.
+                    - `{"max_height": int, "max_width": int}`: The image will be resized to the maximum size respecting the
+                        aspect ratio and keeping the height less or equal to `max_height` and the width less or equal to
+                        `max_width`.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BILINEAR`):
+                Resampling filter to use if resizing the image.
+        """
+        resample = resample if resample is not None else self.resample
+
+        if size.shortest_edge and size.longest_edge:
+            # Resize the image so that the shortest edge or the longest edge is of the given size
+            # while maintaining the aspect ratio of the original image.
+            new_size = get_size_with_aspect_ratio(
+                image.shape[-2:],
+                size.shortest_edge,
+                size.longest_edge or size.shortest_edge,
+            )
+        elif size.max_height and size.max_width:
+            new_size = get_image_size_for_max_height_width(image.shape[-2:], size.max_height, size.max_width)
+        elif size.height and size.width:
+            new_size = (size.height, size.width)
+        else:
+            raise ValueError(
+                f"Size must contain 'height' and 'width' keys or 'shortest_edge' and 'longest_edge' keys. Got {size}."
+            )
+
+        image = super().resize(
+            image,
+            size=SizeDict(height=new_size[0], width=new_size[1]),
+            resample=resample,
+            **kwargs,
+        )
+        return image
+
+    def resize_annotation(
+        self,
+        annotation: dict[str, Any],
+        orig_size: tuple[int, int],
+        target_size: tuple[int, int],
+        threshold: float = 0.5,
+        resample: Optional["PILImageResampling"] = PILImageResampling.NEAREST,
+    ):
+        """
+        Resizes an annotation to a target size.
+
+        Args:
+            annotation (`dict[str, Any]`):
+                The annotation dictionary.
+            orig_size (`tuple[int, int]`):
+                The original size of the input image.
+            target_size (`tuple[int, int]`):
+                The target size of the image, as returned by the preprocessing `resize` step.
+            threshold (`float`, *optional*, defaults to 0.5):
+                The threshold used to binarize the segmentation masks.
+            resample (`PILImageResampling`, defaults to `PILImageResampling.NEAREST`):
+                The resampling filter to use when resizing the masks.
+        """
+        ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(target_size, orig_size))
+        ratio_height, ratio_width = ratios
+
+        new_annotation = {}
+        new_annotation["size"] = target_size
+
+        for key, value in annotation.items():
+            if key == "boxes":
+                boxes = value
+                scaled_boxes = boxes * np.asarray(
+                    [ratio_width, ratio_height, ratio_width, ratio_height], dtype=np.float32
+                )
+                new_annotation["boxes"] = scaled_boxes
+            elif key == "area":
+                area = value
+                scaled_area = area * (ratio_width * ratio_height)
+                new_annotation["area"] = scaled_area
+            elif key == "masks":
+                masks = value[:, None]
+                masks = np.array([resize(mask, target_size, resample=resample) for mask in masks])
+                masks = masks.astype(np.float32)
+                masks = masks[:, 0] > threshold
+                new_annotation["masks"] = masks
+            elif key == "size":
+                new_annotation["size"] = target_size
+            else:
+                new_annotation[key] = value
+
+        return new_annotation
+
+    def normalize_annotation(self, annotation: dict, image_size: tuple[int, int]) -> dict:
+        image_height, image_width = image_size
+        norm_annotation = {}
+        for key, value in annotation.items():
+            if key == "boxes":
+                boxes = value
+                boxes = corners_to_center_format(boxes)
+                boxes /= np.asarray([image_width, image_height, image_width, image_height], dtype=np.float32)
+                norm_annotation[key] = boxes
+            else:
+                norm_annotation[key] = value
+        return norm_annotation
+
+    def _update_annotation_for_padded_image(
+        self,
+        annotation: dict,
+        input_image_size: tuple[int, int],
+        output_image_size: tuple[int, int],
+        padding,
+        update_bboxes,
+    ) -> dict:
+        """
+        Update the annotation for a padded image.
+        """
+        new_annotation = {}
+        new_annotation["size"] = output_image_size
+        ratio_height, ratio_width = (input / output for output, input in zip(output_image_size, input_image_size))
+
+        for key, value in annotation.items():
+            if key == "masks":
+                masks = value
+                masks = pad(
+                    masks,
+                    padding,
+                    mode=PaddingMode.CONSTANT,
+                    constant_values=0,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+                masks = safe_squeeze(masks, 1)
+                new_annotation["masks"] = masks
+            elif key == "boxes" and update_bboxes:
+                boxes = value
+                boxes *= np.asarray(
+                    [
+                        input_image_size[1] / output_image_size[1],
+                        input_image_size[0] / output_image_size[0],
+                        input_image_size[1] / output_image_size[1],
+                        input_image_size[0] / output_image_size[0],
+                    ]
+                )
+                new_annotation["boxes"] = boxes
+            elif key == "size":
+                new_annotation["size"] = output_image_size
+            else:
+                new_annotation[key] = value
+        return new_annotation
+
+    def pad(
+        self,
+        image: np.ndarray,
+        padded_size: tuple[int, int],
+        annotation: dict[str, Any] | None = None,
+        update_bboxes: bool = True,
+        fill: int = 0,
+    ):
+        input_height, input_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        output_height, output_width = padded_size
+        padding_bottom = output_height - input_height
+        padding_right = output_width - input_width
+        if padding_bottom < 0 or padding_right < 0:
+            raise ValueError(
+                f"Padding dimensions are negative. Please make sure that the padded size is larger than the "
+                f"original size. Got padded size: {padded_size}, original size: {(input_height, input_width)}."
+            )
+        if (input_height, input_width) != padded_size:
+            padding = ((0, padding_bottom), (0, padding_right))
+            image = pad(
+                image,
+                padding,
+                mode=PaddingMode.CONSTANT,
+                constant_values=fill,
+                data_format=ChannelDimension.FIRST,
+                input_data_format=ChannelDimension.FIRST,
+            )
+            if annotation is not None:
+                annotation = self._update_annotation_for_padded_image(
+                    annotation, (input_height, input_width), (output_height, output_width), padding, update_bboxes
+                )
+
+        # Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
+        pixel_mask = np.zeros(padded_size, dtype=np.int64)
+        pixel_mask[:input_height, :input_width] = 1
+
+        return image, pixel_mask, annotation
+
+    @auto_docstring
+    def preprocess(
+        self,
+        images: ImageInput,
+        annotations: AnnotationType | list[AnnotationType] | None = None,
+        return_segmentation_masks: bool | None = None,
+        masks_path: str | pathlib.Path | None = None,
+        **kwargs: Unpack[DeformableDetrImageProcessorKwargs],
+    ) -> BatchFeature:
+        r"""
+        annotations (`AnnotationType` or `list[AnnotationType]`, *optional*):
+            Annotations to transform according to the padding that is applied to the images.
+        return_segmentation_masks (`bool`, *optional*, defaults to `self.return_segmentation_masks`):
+            Whether to return segmentation masks.
+        masks_path (`str` or `pathlib.Path`, *optional*):
+            Path to the directory containing the segmentation masks.
+        """
+        return super().preprocess(images, annotations, return_segmentation_masks, masks_path, **kwargs)
+
+    def _preprocess(
+        self,
+        images: list[np.ndarray],
+        annotations: AnnotationType | list[AnnotationType] | None,
+        return_segmentation_masks: bool,
+        masks_path: str | pathlib.Path | None,
+        do_resize: bool,
+        size: SizeDict,
+        resample: "PILImageResampling | None",
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        do_convert_annotations: bool,
+        image_mean: float | list[float] | None,
+        image_std: float | list[float] | None,
+        do_pad: bool,
+        pad_size: SizeDict | None,
+        format: str | AnnotationFormat | None,
+        return_tensors: str | TensorType | None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Preprocess an image or a batch of images so that it can be used by the model.
+        """
+        if annotations is not None and isinstance(annotations, dict):
+            annotations = [annotations]
+
+        if annotations is not None and len(images) != len(annotations):
+            raise ValueError(
+                f"The number of images ({len(images)}) and annotations ({len(annotations)}) do not match."
+            )
+
+        format = AnnotationFormat(format)
+        if annotations is not None:
+            validate_annotations(format, SUPPORTED_ANNOTATION_FORMATS, annotations)
+
+        if (
+            masks_path is not None
+            and format == AnnotationFormat.COCO_PANOPTIC
+            and not isinstance(masks_path, (pathlib.Path, str))
+        ):
+            raise ValueError(
+                "The path to the directory containing the mask PNG files should be provided as a"
+                f" `pathlib.Path` or string object, but is {type(masks_path)} instead."
+            )
+
+        data = {}
+
+        # Import torch if needed for tensor conversion
+        if return_tensors == "pt":
+            if not is_torch_available():
+                raise ImportError("PyTorch is required for tensor conversion.")
+
+        processed_images = []
+        processed_annotations = []
+        pixel_masks = []  # Initialize pixel_masks here
+        for image, annotation in zip(images, annotations if annotations is not None else [None] * len(images)):
+            # prepare (COCO annotations as a list of Dict -> DEFORMABLE_DETR target as a single Dict per image)
+            if annotations is not None:
+                annotation = self.prepare_annotation(
+                    image,
+                    annotation,
+                    format,
+                    return_segmentation_masks=return_segmentation_masks,
+                    masks_path=masks_path,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+
+            if do_resize:
+                resized_image = self.resize(image, size=size, resample=resample)
+                if annotations is not None:
+                    annotation = self.resize_annotation(
+                        annotation,
+                        orig_size=get_image_size(image, channel_dim=ChannelDimension.FIRST),
+                        target_size=get_image_size(resized_image, channel_dim=ChannelDimension.FIRST),
+                    )
+                image = resized_image
+
+            if do_rescale:
+                image = self.rescale(image, rescale_factor)
+            if do_normalize:
+                image = self.normalize(image, image_mean, image_std)
+
+            if do_convert_annotations and annotations is not None:
+                annotation = self.normalize_annotation(annotation, get_image_size(image, ChannelDimension.FIRST))
+
+            processed_images.append(image)
+            processed_annotations.append(annotation)
+        images = processed_images
+        annotations = processed_annotations if annotations is not None else None
+
+        if do_pad:
+            # depends on all resized image shapes so we need another loop
+            if pad_size is not None:
+                padded_size = (pad_size.height, pad_size.width)
+            else:
+                padded_size = get_max_height_width(images, input_data_format=ChannelDimension.FIRST)
+
+            padded_images = []
+            padded_annotations = []
+            for image, annotation in zip(images, annotations if annotations is not None else [None] * len(images)):
+                # Pads images and returns their mask: {'pixel_values': ..., 'pixel_mask': ...}
+                image_height, image_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+                if padded_size == (image_height, image_width):
+                    padded_images.append(image)
+                    pixel_masks.append(np.ones(padded_size, dtype=np.int64))
+                    padded_annotations.append(annotation)
+                    continue
+                image, pixel_mask, annotation = self.pad(
+                    image, padded_size, annotation=annotation, update_bboxes=do_convert_annotations
+                )
+                padded_images.append(image)
+                padded_annotations.append(annotation)
+                pixel_masks.append(pixel_mask)
+            images = padded_images
+            annotations = padded_annotations if annotations is not None else None
+            data.update({"pixel_mask": pixel_masks})
+
+        data.update({"pixel_values": images})
+        encoded_inputs = BatchFeature(data, tensor_type=return_tensors)
+        if annotations is not None:
+            encoded_inputs["labels"] = [
+                BatchFeature(annotation, tensor_type=return_tensors) for annotation in annotations
+            ]
+        return encoded_inputs
+
+    @requires(backends=("torch",))
+    def post_process_object_detection(
+        self, outputs, threshold: float = 0.5, target_sizes: TensorType | list[tuple] = None, top_k: int = 100
+    ):
+        """
+        Converts the raw output of [`DeformableDetrForObjectDetection`] into final bounding boxes in (top_left_x,
+        top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.
+
+        Args:
+            outputs ([`DetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*):
+                Score threshold to keep object detection predictions.
+            target_sizes (`torch.Tensor` or `list[tuple[int, int]]`, *optional*):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`tuple[int, int]`) containing the target size
+                (height, width) of each image in the batch. If left to None, predictions will not be resized.
+            top_k (`int`, *optional*, defaults to 100):
+                Keep only top k bounding boxes before filtering by thresholding.
+
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        requires_backends(self, ["torch"])
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if target_sizes is not None:
+            if len(out_logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+        prob = out_logits.sigmoid()
+        prob = prob.view(out_logits.shape[0], -1)
+        k_value = min(top_k, prob.size(1))
+        topk_values, topk_indexes = torch.topk(prob, k_value, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            if isinstance(target_sizes, list):
+                img_h = torch.Tensor([i[0] for i in target_sizes])
+                img_w = torch.Tensor([i[1] for i in target_sizes])
+            else:
+                img_h, img_w = target_sizes.unbind(1)
+            scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1).to(boxes.device)
+            boxes = boxes * scale_fct[:, None, :]
+
+        results = []
+        for s, l, b in zip(scores, labels, boxes):
+            score = s[s > threshold]
+            label = l[s > threshold]
+            box = b[s > threshold]
+            results.append({"scores": score, "labels": label, "boxes": box})
+
+        return results
+
+
+__all__ = ["DeformableDetrImageProcessorPil"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -0,0 +1,1711 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deformable_detr/modular_deformable_detr.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_deformable_detr.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2022 SenseTime and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+import warnings
+from collections.abc import Callable
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...backbone_utils import load_backbone
+from ...integrations import use_kernel_forward_from_hub
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithCrossAttentions
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...pytorch_utils import compile_compatible_method_lru_cache
+from ...utils import ModelOutput, TransformersKwargs, auto_docstring, torch_compilable_check
+from ...utils.generic import can_return_tuple, merge_with_config_defaults
+from ...utils.output_capturing import OutputRecorder, capture_outputs
+from .configuration_deformable_detr import DeformableDetrConfig
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for outputs of the DEFORMABLE_DETR decoder. This class adds one attribute to BaseModelOutputWithCrossAttentions,
+    namely an optional stack of intermediate decoder activations, i.e. the output of each decoder layer, each of them
+    gone through a layernorm. This is useful when training the model with auxiliary decoding losses.
+    """
+)
+@dataclass
+class DeformableDetrDecoderOutput(BaseModelOutputWithCrossAttentions):
+    r"""
+    cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=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 of the decoder's cross-attention layer, after the attention softmax,
+        used to compute the weighted average in the cross-attention heads.
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(config.decoder_layers, batch_size, num_queries, hidden_size)`, *optional*, returned when `config.auxiliary_loss=True`):
+        Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a
+        layernorm.
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, sequence_length, hidden_size)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    """
+
+    intermediate_hidden_states: torch.FloatTensor | None = None
+
+    intermediate_reference_points: torch.FloatTensor | None = None
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for outputs of the Deformable DETR encoder-decoder model.
+    """
+)
+@dataclass
+class DeformableDetrModelOutput(ModelOutput):
+    r"""
+    init_reference_points (`torch.FloatTensor` of shape  `(batch_size, num_queries, 4)`):
+        Initial reference points sent through the Transformer decoder.
+    last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+        Sequence of hidden-states at the output of the last layer of the decoder of the model.
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+        Stacked intermediate hidden states (output of each layer of the decoder).
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+        picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+        foreground and background).
+    enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the first stage.
+    """
+
+    init_reference_points: torch.FloatTensor | None = None
+    last_hidden_state: torch.FloatTensor | None = None
+    intermediate_hidden_states: torch.FloatTensor | None = None
+    intermediate_reference_points: torch.FloatTensor | None = None
+    decoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    decoder_attentions: tuple[torch.FloatTensor] | None = None
+    cross_attentions: tuple[torch.FloatTensor] | None = None
+    encoder_last_hidden_state: torch.FloatTensor | None = None
+    encoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    encoder_attentions: tuple[torch.FloatTensor] | None = None
+    enc_outputs_class: torch.FloatTensor | None = None
+    enc_outputs_coord_logits: torch.FloatTensor | None = None
+
+
+@auto_docstring(
+    custom_intro="""
+    Output type of [`DeformableDetrForObjectDetection`].
+    """
+)
+@dataclass
+class DeformableDetrObjectDetectionOutput(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` are provided)):
+        Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a
+        bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized
+        scale-invariant IoU loss.
+    loss_dict (`Dict`, *optional*):
+        A dictionary containing the individual losses. Useful for logging.
+    logits (`torch.FloatTensor` of shape `(batch_size, num_queries, num_classes + 1)`):
+        Classification logits (including no-object) for all queries.
+    pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
+        Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
+        values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
+        possible padding). You can use [`~DeformableDetrProcessor.post_process_object_detection`] to retrieve the
+        unnormalized bounding boxes.
+    auxiliary_outputs (`list[Dict]`, *optional*):
+        Optional, only returned when auxiliary losses are activated (i.e. `config.auxiliary_loss` is set to `True`)
+        and labels are provided. It is a list of dictionaries containing the two above keys (`logits` and
+        `pred_boxes`) for each decoder layer.
+    last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+        Sequence of hidden-states at the output of the last layer of the decoder of the model.
+    init_reference_points (`torch.FloatTensor` of shape  `(batch_size, num_queries, 4)`):
+        Initial reference points sent through the Transformer decoder.
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+        Stacked intermediate hidden states (output of each layer of the decoder).
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+        picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+        foreground and background).
+    enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the first stage.
+    """
+
+    loss: torch.FloatTensor | None = None
+    loss_dict: dict | None = None
+    logits: torch.FloatTensor | None = None
+    pred_boxes: torch.FloatTensor | None = None
+    auxiliary_outputs: list[dict] | None = None
+    last_hidden_state: torch.FloatTensor | None = None
+    decoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    decoder_attentions: tuple[torch.FloatTensor] | None = None
+    cross_attentions: tuple[torch.FloatTensor] | None = None
+    encoder_last_hidden_state: torch.FloatTensor | None = None
+    encoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    encoder_attentions: tuple[torch.FloatTensor] | None = None
+
+    init_reference_points: torch.FloatTensor | None = None
+    intermediate_hidden_states: torch.FloatTensor | None = None
+    intermediate_reference_points: torch.FloatTensor | None = None
+    enc_outputs_class: torch.FloatTensor | None = None
+    enc_outputs_coord_logits: torch.FloatTensor | None = None
+
+
+@use_kernel_forward_from_hub("MultiScaleDeformableAttention")
+class MultiScaleDeformableAttention(nn.Module):
+    def forward(
+        self,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_spatial_shapes_list: list[tuple],
+        level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        batch_size, _, num_heads, hidden_dim = value.shape
+        _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
+        value_list = value.split([height * width for height, width in value_spatial_shapes_list], dim=1)
+        sampling_grids = 2 * sampling_locations - 1
+        sampling_value_list = []
+        for level_id, (height, width) in enumerate(value_spatial_shapes_list):
+            # batch_size, height*width, num_heads, hidden_dim
+            # -> batch_size, height*width, num_heads*hidden_dim
+            # -> batch_size, num_heads*hidden_dim, height*width
+            # -> batch_size*num_heads, hidden_dim, height, width
+            value_l_ = (
+                value_list[level_id]
+                .flatten(2)
+                .transpose(1, 2)
+                .reshape(batch_size * num_heads, hidden_dim, height, width)
+            )
+            # batch_size, num_queries, num_heads, num_points, 2
+            # -> batch_size, num_heads, num_queries, num_points, 2
+            # -> batch_size*num_heads, num_queries, num_points, 2
+            sampling_grid_l_ = sampling_grids[:, :, :, level_id].transpose(1, 2).flatten(0, 1)
+            # batch_size*num_heads, hidden_dim, num_queries, num_points
+            sampling_value_l_ = nn.functional.grid_sample(
+                value_l_,
+                sampling_grid_l_,
+                mode="bilinear",
+                padding_mode="zeros",
+                align_corners=False,
+            )
+            sampling_value_list.append(sampling_value_l_)
+        # (batch_size, num_queries, num_heads, num_levels, num_points)
+        # -> (batch_size, num_heads, num_queries, num_levels, num_points)
+        # -> (batch_size, num_heads, 1, num_queries, num_levels*num_points)
+        attention_weights = attention_weights.transpose(1, 2).reshape(
+            batch_size * num_heads, 1, num_queries, num_levels * num_points
+        )
+        output = (
+            (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
+            .sum(-1)
+            .view(batch_size, num_heads * hidden_dim, num_queries)
+        )
+        return output.transpose(1, 2).contiguous()
+
+
+class DeformableDetrFrozenBatchNorm2d(nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than
+    torchvision.models.resnet[18,34,50,101] produce nans.
+    """
+
+    def __init__(self, n):
+        super().__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        num_batches_tracked_key = prefix + "num_batches_tracked"
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super()._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+        )
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it user-friendly
+        weight = self.weight.reshape(1, -1, 1, 1)
+        bias = self.bias.reshape(1, -1, 1, 1)
+        running_var = self.running_var.reshape(1, -1, 1, 1)
+        running_mean = self.running_mean.reshape(1, -1, 1, 1)
+        epsilon = 1e-5
+        scale = weight * (running_var + epsilon).rsqrt()
+        bias = bias - running_mean * scale
+        return x * scale + bias
+
+
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `DeformableDetrFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = DeformableDetrFrozenBatchNorm2d(module.num_features)
+
+            if module.weight.device != torch.device("meta"):
+                new_module.weight.copy_(module.weight)
+                new_module.bias.copy_(module.bias)
+                new_module.running_mean.copy_(module.running_mean)
+                new_module.running_var.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
+
+
+class DeformableDetrConvEncoder(nn.Module):
+    """
+    Convolutional backbone, using either the AutoBackbone API or one from the timm library.
+
+    nn.BatchNorm2d layers are replaced by DeformableDetrFrozenBatchNorm2d as defined above.
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+
+        backbone = load_backbone(config)
+        self.intermediate_channel_sizes = backbone.channels
+
+        # replace batch norm by frozen batch norm
+        with torch.no_grad():
+            replace_batch_norm(backbone)
+
+        # We used to load with timm library directly instead of the AutoBackbone API
+        # so we need to unwrap the `backbone._backbone` module to load weights without mismatch
+        is_timm_model = False
+        if hasattr(backbone, "_backbone"):
+            backbone = backbone._backbone
+            is_timm_model = True
+        self.model = backbone
+
+        backbone_model_type = config.backbone_config.model_type
+        if "resnet" in backbone_model_type:
+            for name, parameter in self.model.named_parameters():
+                if is_timm_model:
+                    if "layer2" not in name and "layer3" not in name and "layer4" not in name:
+                        parameter.requires_grad_(False)
+                else:
+                    if "stage.1" not in name and "stage.2" not in name and "stage.3" not in name:
+                        parameter.requires_grad_(False)
+
+    def forward(self, pixel_values: torch.Tensor, pixel_mask: torch.Tensor):
+        # send pixel_values through the model to get list of feature maps
+        features = self.model(pixel_values)
+        if isinstance(features, dict):
+            features = features.feature_maps
+
+        out = []
+        for feature_map in features:
+            # downsample pixel_mask to match shape of corresponding feature_map
+            mask = nn.functional.interpolate(pixel_mask[None].float(), size=feature_map.shape[-2:]).to(torch.bool)[0]
+            out.append((feature_map, mask))
+        return out
+
+
+class DeformableDetrSinePositionEmbedding(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one used by the Attention is all you
+    need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_position_features: int = 64,
+        temperature: int = 10000,
+        normalize: bool = False,
+        scale: float | None = None,
+    ):
+        super().__init__()
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        self.num_position_features = num_position_features
+        self.temperature = temperature
+        self.normalize = normalize
+        self.scale = 2 * math.pi if scale is None else scale
+
+    @staticmethod
+    @compile_compatible_method_lru_cache(maxsize=1)
+    def build_sine_position_embedding(
+        shape: torch.Size,
+        device: torch.device | str,
+        dtype: torch.dtype,
+        num_position_features: int,
+        normalize: bool = False,
+        scale: float | None = None,
+        temperature: int = 10000,
+        mask: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        if mask is None:
+            mask = torch.ones((shape[0], shape[2], shape[3]), device=device, dtype=torch.bool)
+        y_embed = mask.cumsum(1, dtype=dtype)
+        x_embed = mask.cumsum(2, dtype=dtype)
+        if normalize:
+            eps = 1e-6
+            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * scale
+            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * scale
+
+        dim_t = torch.arange(num_position_features, dtype=torch.int64, device=device).to(dtype)
+        dim_t = temperature ** (2 * torch.div(dim_t, 2, rounding_mode="floor") / num_position_features)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+
+    def forward(
+        self,
+        shape: torch.Size,
+        device: torch.device | str,
+        dtype: torch.dtype,
+        mask: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        return self.build_sine_position_embedding(
+            shape, device, dtype, self.num_position_features, self.normalize, self.scale, self.temperature, mask
+        )
+
+
+class DeformableDetrLearnedPositionEmbedding(nn.Module):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, embedding_dim=256):
+        super().__init__()
+        self.row_embeddings = nn.Embedding(50, embedding_dim)
+        self.column_embeddings = nn.Embedding(50, embedding_dim)
+
+    @compile_compatible_method_lru_cache(maxsize=1)
+    def forward(
+        self,
+        shape: torch.Size,
+        device: torch.device | str,
+        dtype: torch.dtype,
+        mask: torch.Tensor | None = None,
+    ):
+        height, width = shape[-2:]
+        width_values = torch.arange(width, device=device)
+        height_values = torch.arange(height, device=device)
+        x_emb = self.column_embeddings(width_values)
+        y_emb = self.row_embeddings(height_values)
+        pos = torch.cat([x_emb.unsqueeze(0).repeat(height, 1, 1), y_emb.unsqueeze(1).repeat(1, width, 1)], dim=-1)
+        pos = pos.permute(2, 0, 1)
+        pos = pos.unsqueeze(0)
+        pos = pos.repeat(shape[0], 1, 1, 1)
+        return pos
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float | None = None,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    if scaling is None:
+        scaling = query.size(-1) ** -0.5
+
+    # Take the dot product between "query" and "key" to get the raw attention scores.
+    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
+
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+
+    attn_output = torch.matmul(attn_weights, value)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class DeformableDetrSelfAttention(nn.Module):
+    """
+    Multi-headed self-attention from 'Attention Is All You Need' paper.
+
+    In DEFORMABLE_DETR, position embeddings are added to both queries and keys (but not values) in self-attention.
+    """
+
+    def __init__(
+        self,
+        config: DeformableDetrConfig,
+        hidden_size: int,
+        num_attention_heads: int,
+        dropout: float = 0.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.config = config
+        self.head_dim = hidden_size // num_attention_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = dropout
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.v_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.q_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.o_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_embeddings: torch.Tensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Position embeddings are added to both queries and keys (but not values).
+        """
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_key_input = hidden_states + position_embeddings if position_embeddings is not None else hidden_states
+
+        query_states = self.q_proj(query_key_input).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(query_key_input).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class DeformableDetrMultiscaleDeformableAttention(nn.Module):
+    """
+    Multiscale deformable attention as proposed in Deformable DETR.
+    """
+
+    def __init__(self, config: DeformableDetrConfig, num_heads: int, n_points: int):
+        super().__init__()
+
+        self.attn = MultiScaleDeformableAttention()
+
+        if config.d_model % num_heads != 0:
+            raise ValueError(
+                f"embed_dim (d_model) must be divisible by num_heads, but got {config.d_model} and {num_heads}"
+            )
+        dim_per_head = config.d_model // num_heads
+        # check if dim_per_head is power of 2
+        if not ((dim_per_head & (dim_per_head - 1) == 0) and dim_per_head != 0):
+            warnings.warn(
+                "You'd better set embed_dim (d_model) in DeformableDetrMultiscaleDeformableAttention to make the"
+                " dimension of each attention head a power of 2 which is more efficient in the authors' CUDA"
+                " implementation."
+            )
+
+        self.im2col_step = 64
+
+        self.d_model = config.d_model
+        self.n_levels = config.num_feature_levels
+        self.n_heads = num_heads
+        self.n_points = n_points
+
+        self.sampling_offsets = nn.Linear(config.d_model, num_heads * self.n_levels * n_points * 2)
+        self.attention_weights = nn.Linear(config.d_model, num_heads * self.n_levels * n_points)
+        self.value_proj = nn.Linear(config.d_model, config.d_model)
+        self.output_proj = nn.Linear(config.d_model, config.d_model)
+
+        self.disable_custom_kernels = config.disable_custom_kernels
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        position_embeddings: torch.Tensor | None = None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # add position embeddings to the hidden states before projecting to queries and keys
+        if position_embeddings is not None:
+            hidden_states = hidden_states + position_embeddings
+
+        batch_size, num_queries, _ = hidden_states.shape
+        batch_size, sequence_length, _ = encoder_hidden_states.shape
+        total_elements = sum(height * width for height, width in spatial_shapes_list)
+        torch_compilable_check(
+            total_elements == sequence_length,
+            "Make sure to align the spatial shapes with the sequence length of the encoder hidden states",
+        )
+
+        value = self.value_proj(encoder_hidden_states)
+        if attention_mask is not None:
+            # we invert the attention_mask
+            value = value.masked_fill(~attention_mask[..., None], float(0))
+        value = value.view(batch_size, sequence_length, self.n_heads, self.d_model // self.n_heads)
+        sampling_offsets = self.sampling_offsets(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points, 2
+        )
+        attention_weights = self.attention_weights(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels * self.n_points
+        )
+        attention_weights = F.softmax(attention_weights, -1).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points
+        )
+        # batch_size, num_queries, n_heads, n_levels, n_points, 2
+        num_coordinates = reference_points.shape[-1]
+        if num_coordinates == 2:
+            offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif num_coordinates == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2]
+                + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5
+            )
+        else:
+            raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
+
+        output = self.attn(
+            value,
+            spatial_shapes,
+            spatial_shapes_list,
+            level_start_index,
+            sampling_locations,
+            attention_weights,
+            self.im2col_step,
+        )
+
+        output = self.output_proj(output)
+
+        return output, attention_weights
+
+
+class DeformableDetrMLP(nn.Module):
+    def __init__(self, config: DeformableDetrConfig, hidden_size: int, intermediate_size: int):
+        super().__init__()
+        self.fc1 = nn.Linear(hidden_size, intermediate_size)
+        self.fc2 = nn.Linear(intermediate_size, hidden_size)
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.dropout = config.dropout
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        return hidden_states
+
+
+class DeformableDetrEncoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__()
+        self.hidden_size = config.d_model
+        self.self_attn = DeformableDetrMultiscaleDeformableAttention(
+            config,
+            num_heads=config.encoder_attention_heads,
+            n_points=config.encoder_n_points,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.hidden_size)
+        self.dropout = config.dropout
+        self.mlp = DeformableDetrMLP(config, self.hidden_size, config.encoder_ffn_dim)
+        self.final_layer_norm = nn.LayerNorm(self.hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        spatial_position_embeddings: torch.Tensor | None = None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Input to the layer.
+            attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
+                Attention mask.
+            position_embeddings (`torch.FloatTensor`, *optional*):
+                Position embeddings, to be added to `hidden_states`.
+            reference_points (`torch.FloatTensor`, *optional*):
+                Reference points.
+            spatial_shapes (`torch.LongTensor`, *optional*):
+                Spatial shapes of the backbone feature maps.
+            level_start_index (`torch.LongTensor`, *optional*):
+                Level start index.
+        """
+        residual = hidden_states
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            position_embeddings=spatial_position_embeddings,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
+            level_start_index=level_start_index,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.training:
+            if not torch.isfinite(hidden_states).all():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        return hidden_states
+
+
+class DeformableDetrDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__()
+        self.hidden_size = config.d_model
+
+        self.self_attn = DeformableDetrSelfAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_attention_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.dropout = config.dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.hidden_size)
+        self.encoder_attn = DeformableDetrMultiscaleDeformableAttention(
+            config,
+            num_heads=config.decoder_attention_heads,
+            n_points=config.decoder_n_points,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.hidden_size)
+        self.mlp = DeformableDetrMLP(config, self.hidden_size, config.decoder_ffn_dim)
+        self.final_layer_norm = nn.LayerNorm(self.hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        object_queries_position_embeddings: torch.Tensor | None = None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        encoder_hidden_states: torch.Tensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(seq_len, batch, embed_dim)`.
+            position_embeddings (`torch.FloatTensor`, *optional*):
+                Position embeddings that are added to the queries and keys in the self-attention layer.
+            reference_points (`torch.FloatTensor`, *optional*):
+                Reference points.
+            spatial_shapes (`torch.LongTensor`, *optional*):
+                Spatial shapes.
+            level_start_index (`torch.LongTensor`, *optional*):
+                Level start index.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, target_len, source_len)` where padding elements are indicated by very large negative
+                values.
+        """
+        residual = hidden_states
+
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=object_queries_position_embeddings,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+
+        # Cross-Attention
+        hidden_states, _ = self.encoder_attn(
+            hidden_states=hidden_states,
+            attention_mask=encoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            position_embeddings=object_queries_position_embeddings,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
+            level_start_index=level_start_index,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return hidden_states
+
+
+@auto_docstring
+class DeformableDetrPreTrainedModel(PreTrainedModel):
+    config: DeformableDetrConfig
+    base_model_prefix = "model"
+    main_input_name = "pixel_values"
+    input_modalities = ("image",)
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        r"DeformableDetrConvEncoder",
+        r"DeformableDetrEncoderLayer",
+        r"DeformableDetrDecoderLayer",
+    ]
+    _supports_sdpa = True
+    _supports_flash_attn = True
+    _supports_attention_backend = True
+    _supports_flex_attn = True
+    _keys_to_ignore_on_load_unexpected = [
+        r"detr\.model\.backbone\.model\.layer\d+\.0\.downsample\.1\.num_batches_tracked"
+    ]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        std = self.config.init_std
+
+        if isinstance(module, DeformableDetrLearnedPositionEmbedding):
+            init.uniform_(module.row_embeddings.weight)
+            init.uniform_(module.column_embeddings.weight)
+        elif isinstance(module, DeformableDetrMultiscaleDeformableAttention):
+            init.constant_(module.sampling_offsets.weight, 0.0)
+            default_dtype = torch.get_default_dtype()
+            thetas = torch.arange(module.n_heads, dtype=torch.int64).to(default_dtype) * (
+                2.0 * math.pi / module.n_heads
+            )
+            grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+            grid_init = (
+                (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+                .view(module.n_heads, 1, 1, 2)
+                .repeat(1, module.n_levels, module.n_points, 1)
+            )
+            for i in range(module.n_points):
+                grid_init[:, :, i, :] *= i + 1
+
+            init.copy_(module.sampling_offsets.bias, grid_init.view(-1))
+            init.constant_(module.attention_weights.weight, 0.0)
+            init.constant_(module.attention_weights.bias, 0.0)
+            init.xavier_uniform_(module.value_proj.weight)
+            init.constant_(module.value_proj.bias, 0.0)
+            init.xavier_uniform_(module.output_proj.weight)
+            init.constant_(module.output_proj.bias, 0.0)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            init.normal_(module.weight, mean=0.0, std=std)
+            if module.bias is not None:
+                init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            init.normal_(module.weight, mean=0.0, std=std)
+            # Here we need the check explicitly, as we slice the weight in the `zeros_` call, so it looses the flag
+            if module.padding_idx is not None and not getattr(module.weight, "_is_hf_initialized", False):
+                init.zeros_(module.weight[module.padding_idx])
+        if hasattr(module, "reference_points") and not self.config.two_stage:
+            init.xavier_uniform_(module.reference_points.weight, gain=1.0)
+            init.constant_(module.reference_points.bias, 0.0)
+        if hasattr(module, "level_embed"):
+            init.normal_(module.level_embed)
+
+
+class DeformableDetrEncoder(DeformableDetrPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* deformable attention layers. Each layer is a
+    [`DeformableDetrEncoderLayer`].
+
+    The encoder updates the flattened multi-scale feature maps through multiple deformable attention layers.
+
+    Args:
+        config: DeformableDetrConfig
+    """
+
+    _can_record_outputs = {
+        "hidden_states": DeformableDetrEncoderLayer,
+        "attentions": OutputRecorder(DeformableDetrMultiscaleDeformableAttention, layer_name="self_attn", index=1),
+    }
+
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layers = nn.ModuleList([DeformableDetrEncoderLayer(config) for _ in range(config.encoder_layers)])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @merge_with_config_defaults
+    @capture_outputs
+    def forward(
+        self,
+        inputs_embeds=None,
+        attention_mask=None,
+        spatial_position_embeddings=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        valid_ratios=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutput:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Flattened feature map (output of the backbone + projection layer) that is passed to the encoder.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding pixel features. Mask values selected in `[0, 1]`:
+                - 1 for pixel features that are real (i.e. **not masked**),
+                - 0 for pixel features that are padding (i.e. **masked**).
+                [What are attention masks?](../glossary#attention-mask)
+            spatial_position_embeddings (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Spatial position embeddings (2D positional encodings) that are added to the queries and keys in each self-attention layer.
+            spatial_shapes (`torch.LongTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of each feature map.
+            level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`):
+                Starting index of each feature map.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`):
+                Ratio of valid area in each feature level.
+        """
+        hidden_states = inputs_embeds
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        spatial_shapes_tuple = tuple(spatial_shapes_list)
+        reference_points = self.get_reference_points(spatial_shapes_tuple, valid_ratios, device=inputs_embeds.device)
+
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(
+                hidden_states,
+                attention_mask,
+                spatial_position_embeddings=spatial_position_embeddings,
+                reference_points=reference_points,
+                spatial_shapes=spatial_shapes,
+                spatial_shapes_list=spatial_shapes_list,
+                level_start_index=level_start_index,
+                **kwargs,
+            )
+
+        return BaseModelOutput(last_hidden_state=hidden_states)
+
+    @staticmethod
+    def get_reference_points(spatial_shapes_list, valid_ratios, device):
+        """
+        Get reference points for each feature map. Used in decoder.
+
+        Args:
+            spatial_shapes_list (`list[tuple[int, int]]`):
+                Spatial shapes of each feature map.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`):
+                Valid ratios of each feature map.
+            device (`torch.device`):
+                Device on which to create the tensors.
+        Returns:
+            `torch.FloatTensor` of shape `(batch_size, num_queries, num_feature_levels, 2)`
+        """
+        reference_points_list = []
+        for level, (height, width) in enumerate(spatial_shapes_list):
+            ref_y, ref_x = torch.meshgrid(
+                torch.linspace(0.5, height - 0.5, height, dtype=valid_ratios.dtype, device=device),
+                torch.linspace(0.5, width - 0.5, width, dtype=valid_ratios.dtype, device=device),
+                indexing="ij",
+            )
+            # TODO: valid_ratios could be useless here. check https://github.com/fundamentalvision/Deformable-DETR/issues/36
+            ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, level, 1] * height)
+            ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, level, 0] * width)
+            ref = torch.stack((ref_x, ref_y), -1)
+            reference_points_list.append(ref)
+        reference_points = torch.cat(reference_points_list, 1)
+        reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+        return reference_points
+
+
+def inverse_sigmoid(x, eps=1e-5):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+class DeformableDetrDecoder(DeformableDetrPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`DeformableDetrDecoderLayer`].
+
+    The decoder updates the query embeddings through multiple self-attention and cross-attention layers.
+
+    Some tweaks for Deformable DETR:
+
+    - `position_embeddings`, `reference_points`, `spatial_shapes` and `valid_ratios` are added to the forward pass.
+    - it also returns a stack of intermediate outputs and reference points from all decoding layers.
+
+    Args:
+        config: DeformableDetrConfig
+    """
+
+    _can_record_outputs = {
+        "hidden_states": DeformableDetrDecoderLayer,
+        "attentions": OutputRecorder(DeformableDetrSelfAttention, layer_name="self_attn", index=1),
+        "cross_attentions": OutputRecorder(
+            DeformableDetrMultiscaleDeformableAttention, layer_name="encoder_attn", index=1
+        ),
+    }
+
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layers = nn.ModuleList([DeformableDetrDecoderLayer(config) for _ in range(config.decoder_layers)])
+
+        # hack implementation for iterative bounding box refinement and two-stage Deformable DETR
+        self.bbox_embed = None
+        self.class_embed = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @merge_with_config_defaults
+    @capture_outputs
+    def forward(
+        self,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        object_queries_position_embeddings=None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        valid_ratios=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ):
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+                The query embeddings that are passed into the decoder.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding pixel_values of the encoder. Mask values selected
+                in `[0, 1]`:
+                - 1 for pixels that are real (i.e. **not masked**),
+                - 0 for pixels that are padding (i.e. **masked**).
+            object_queries_position_embeddings (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+                Position embeddings for the object query slots that are added to the queries and keys in each self-attention layer.
+            reference_points (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)` is `as_two_stage` else `(batch_size, num_queries, 2)` or , *optional*):
+                Reference point in range `[0, 1]`, top-left (0,0), bottom-right (1, 1), including padding area.
+            spatial_shapes (`torch.FloatTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of the feature maps.
+            level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`, *optional*):
+                Indexes for the start of each feature level. In range `[0, sequence_length]`.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`, *optional*):
+                Ratio of valid area in each feature level.
+
+        """
+        if inputs_embeds is not None:
+            hidden_states = inputs_embeds
+
+        # decoder layers
+        intermediate = ()
+        intermediate_reference_points = ()
+
+        for idx, decoder_layer in enumerate(self.layers):
+            num_coordinates = reference_points.shape[-1]
+            if num_coordinates == 4:
+                reference_points_input = (
+                    reference_points[:, :, None] * torch.cat([valid_ratios, valid_ratios], -1)[:, None]
+                )
+            elif reference_points.shape[-1] == 2:
+                reference_points_input = reference_points[:, :, None] * valid_ratios[:, None]
+            else:
+                raise ValueError("Reference points' last dimension must be of size 2")
+
+            hidden_states = decoder_layer(
+                hidden_states,
+                object_queries_position_embeddings,
+                reference_points_input,
+                spatial_shapes,
+                spatial_shapes_list,
+                level_start_index,
+                encoder_hidden_states,  # as a positional argument for gradient checkpointing
+                encoder_attention_mask,
+                **kwargs,
+            )
+
+            # hack implementation for iterative bounding box refinement
+            if self.bbox_embed is not None:
+                tmp = self.bbox_embed[idx](hidden_states)
+                num_coordinates = reference_points.shape[-1]
+                if num_coordinates == 4:
+                    new_reference_points = tmp + inverse_sigmoid(reference_points)
+                    new_reference_points = new_reference_points.sigmoid()
+                elif num_coordinates == 2:
+                    new_reference_points = tmp
+                    new_reference_points[..., :2] = tmp[..., :2] + inverse_sigmoid(reference_points)
+                    new_reference_points = new_reference_points.sigmoid()
+                else:
+                    raise ValueError(
+                        f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}"
+                    )
+                reference_points = new_reference_points.detach()
+
+            intermediate += (hidden_states,)
+            intermediate_reference_points += (reference_points,)
+
+        # Keep batch_size as first dimension
+        intermediate = torch.stack(intermediate, dim=1)
+        intermediate_reference_points = torch.stack(intermediate_reference_points, dim=1)
+
+        return DeformableDetrDecoderOutput(
+            last_hidden_state=hidden_states,
+            intermediate_hidden_states=intermediate,
+            intermediate_reference_points=intermediate_reference_points,
+        )
+
+
+@auto_docstring(
+    custom_intro="""
+    The bare Deformable DETR Model (consisting of a backbone and encoder-decoder Transformer) outputting raw
+    hidden-states without any specific head on top.
+    """
+)
+class DeformableDetrModel(DeformableDetrPreTrainedModel):
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__(config)
+
+        # Create backbone
+        self.backbone = DeformableDetrConvEncoder(config)
+
+        # Create positional encoding
+        if config.position_embedding_type == "sine":
+            self.position_embedding = DeformableDetrSinePositionEmbedding(config.d_model // 2, normalize=True)
+        elif config.position_embedding_type == "learned":
+            self.position_embedding = DeformableDetrLearnedPositionEmbedding(config.d_model // 2)
+        else:
+            raise ValueError(f"Not supported {config.position_embedding_type}")
+
+        # Create input projection layers
+        if config.num_feature_levels > 1:
+            num_backbone_outs = len(self.backbone.intermediate_channel_sizes)
+            input_proj_list = []
+            for _ in range(num_backbone_outs):
+                in_channels = self.backbone.intermediate_channel_sizes[_]
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, config.d_model, kernel_size=1),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                )
+            for _ in range(config.num_feature_levels - num_backbone_outs):
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(
+                            in_channels,
+                            config.d_model,
+                            kernel_size=3,
+                            stride=2,
+                            padding=1,
+                        ),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                )
+                in_channels = config.d_model
+            self.input_proj = nn.ModuleList(input_proj_list)
+        else:
+            self.input_proj = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        nn.Conv2d(
+                            self.backbone.intermediate_channel_sizes[-1],
+                            config.d_model,
+                            kernel_size=1,
+                        ),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                ]
+            )
+
+        if not config.two_stage:
+            self.query_position_embeddings = nn.Embedding(config.num_queries, config.d_model * 2)
+
+        self.encoder = DeformableDetrEncoder(config)
+        self.decoder = DeformableDetrDecoder(config)
+
+        self.level_embed = nn.Parameter(torch.Tensor(config.num_feature_levels, config.d_model))
+
+        if config.two_stage:
+            self.enc_output = nn.Linear(config.d_model, config.d_model)
+            self.enc_output_norm = nn.LayerNorm(config.d_model)
+            self.pos_trans = nn.Linear(config.d_model * 2, config.d_model * 2)
+            self.pos_trans_norm = nn.LayerNorm(config.d_model * 2)
+        else:
+            self.reference_points = nn.Linear(config.d_model, 2)
+
+        self.post_init()
+
+    def freeze_backbone(self):
+        for name, param in self.backbone.model.named_parameters():
+            param.requires_grad_(False)
+
+    def unfreeze_backbone(self):
+        for name, param in self.backbone.model.named_parameters():
+            param.requires_grad_(True)
+
+    def get_valid_ratio(self, mask, dtype=torch.float32):
+        """Get the valid ratio of all feature maps."""
+
+        _, height, width = mask.shape
+        valid_height = torch.sum(mask[:, :, 0], 1)
+        valid_width = torch.sum(mask[:, 0, :], 1)
+        valid_ratio_height = valid_height.to(dtype) / height
+        valid_ratio_width = valid_width.to(dtype) / width
+        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_height], -1)
+        return valid_ratio
+
+    def get_proposal_pos_embed(self, proposals):
+        """Get the position embedding of the proposals."""
+
+        num_pos_feats = self.config.d_model // 2
+        temperature = 10000
+        scale = 2 * math.pi
+
+        # Compute position embeddings in float32 to avoid overflow with large temperature values in fp16
+        proposals_dtype = proposals.dtype
+        dim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=proposals.device)
+        dim_t = temperature ** (2 * torch.div(dim_t, 2, rounding_mode="floor") / num_pos_feats)
+        # batch_size, num_queries, 4
+        proposals = proposals.sigmoid().to(torch.float32) * scale
+        # batch_size, num_queries, 4, 128
+        pos = proposals[:, :, :, None] / dim_t
+        # batch_size, num_queries, 4, 64, 2 -> batch_size, num_queries, 512
+        pos = torch.stack((pos[:, :, :, 0::2].sin(), pos[:, :, :, 1::2].cos()), dim=4).flatten(2)
+        # Convert back to target dtype after all computations are done
+        return pos.to(proposals_dtype)
+
+    def gen_encoder_output_proposals(self, enc_output, padding_mask, spatial_shapes):
+        """Generate the encoder output proposals from encoded enc_output.
+
+        Args:
+            enc_output (Tensor[batch_size, sequence_length, hidden_size]): Output of the encoder.
+            padding_mask (Tensor[batch_size, sequence_length]): Padding mask for `enc_output`.
+            spatial_shapes (list[tuple[int, int]]): Spatial shapes of the feature maps.
+
+        Returns:
+            `tuple(torch.FloatTensor)`: A tuple of feature map and bbox prediction.
+                - object_query (Tensor[batch_size, sequence_length, hidden_size]): Object query features. Later used to
+                  directly predict a bounding box. (without the need of a decoder)
+                - output_proposals (Tensor[batch_size, sequence_length, 4]): Normalized proposals, after an inverse
+                  sigmoid.
+        """
+        batch_size = enc_output.shape[0]
+        proposals = []
+        _cur = 0
+        for level, (height, width) in enumerate(spatial_shapes):
+            mask_flatten_ = padding_mask[:, _cur : (_cur + height * width)].view(batch_size, height, width, 1)
+            valid_height = torch.sum(~mask_flatten_[:, :, 0, 0], 1)
+            valid_width = torch.sum(~mask_flatten_[:, 0, :, 0], 1)
+
+            grid_y, grid_x = torch.meshgrid(
+                torch.linspace(
+                    0,
+                    height - 1,
+                    height,
+                    dtype=enc_output.dtype,
+                    device=enc_output.device,
+                ),
+                torch.linspace(
+                    0,
+                    width - 1,
+                    width,
+                    dtype=enc_output.dtype,
+                    device=enc_output.device,
+                ),
+                indexing="ij",
+            )
+            grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)
+
+            scale = torch.cat([valid_width.unsqueeze(-1), valid_height.unsqueeze(-1)], 1).view(batch_size, 1, 1, 2)
+            grid = (grid.unsqueeze(0).expand(batch_size, -1, -1, -1) + 0.5) / scale
+            width_height = torch.ones_like(grid) * 0.05 * (2.0**level)
+            proposal = torch.cat((grid, width_height), -1).view(batch_size, -1, 4)
+            proposals.append(proposal)
+            _cur += height * width
+        output_proposals = torch.cat(proposals, 1)
+        output_proposals_valid = ((output_proposals > 0.01) & (output_proposals < 0.99)).all(-1, keepdim=True)
+        output_proposals = torch.log(output_proposals / (1 - output_proposals))  # inverse sigmoid
+        output_proposals = output_proposals.masked_fill(padding_mask.unsqueeze(-1), float("inf"))
+        output_proposals = output_proposals.masked_fill(~output_proposals_valid, float("inf"))
+
+        # assign each pixel as an object query
+        object_query = enc_output
+        object_query = object_query.masked_fill(padding_mask.unsqueeze(-1), float(0))
+        object_query = object_query.masked_fill(~output_proposals_valid, float(0))
+        object_query = self.enc_output_norm(self.enc_output(object_query))
+        return object_query, output_proposals
+
+    @auto_docstring
+    @can_return_tuple
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        pixel_mask: torch.LongTensor | None = None,
+        decoder_attention_mask: torch.FloatTensor | None = None,
+        encoder_outputs: torch.FloatTensor | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        decoder_inputs_embeds: torch.FloatTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.FloatTensor] | DeformableDetrModelOutput:
+        r"""
+        decoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, num_queries)`, *optional*):
+            Not used by default. Can be used to mask object queries.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
+            can choose to directly pass a flattened representation of an image.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
+            embedded representation.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, DeformableDetrModel
+        >>> from PIL import Image
+        >>> import httpx
+        >>> from io import BytesIO
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("SenseTime/deformable-detr")
+        >>> model = DeformableDetrModel.from_pretrained("SenseTime/deformable-detr")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 300, 256]
+        ```"""
+        batch_size, num_channels, height, width = pixel_values.shape
+        device = pixel_values.device
+
+        if pixel_mask is None:
+            pixel_mask = torch.ones(((batch_size, height, width)), dtype=torch.long, device=device)
+
+        # Extract multi-scale feature maps of same resolution `config.d_model` (cf Figure 4 in paper)
+        # First, sent pixel_values + pixel_mask through Backbone to obtain the features
+        # which is a list of tuples
+        features = self.backbone(pixel_values, pixel_mask)
+
+        # Then, apply 1x1 convolution to reduce the channel dimension to d_model (256 by default)
+        sources = []
+        masks = []
+        position_embeddings_list = []
+        for level, (source, mask) in enumerate(features):
+            sources.append(self.input_proj[level](source))
+            masks.append(mask)
+            if mask is None:
+                raise ValueError("No attention mask was provided")
+            # Generate position embeddings for this feature level
+            pos = self.position_embedding(shape=source.shape, device=device, dtype=pixel_values.dtype, mask=mask).to(
+                source.dtype
+            )
+            position_embeddings_list.append(pos)
+
+        # Lowest resolution feature maps are obtained via 3x3 stride 2 convolutions on the final stage
+        if self.config.num_feature_levels > len(sources):
+            _len_sources = len(sources)
+            for level in range(_len_sources, self.config.num_feature_levels):
+                if level == _len_sources:
+                    source = self.input_proj[level](features[-1][0])
+                else:
+                    source = self.input_proj[level](sources[-1])
+                mask = nn.functional.interpolate(pixel_mask[None].to(pixel_values.dtype), size=source.shape[-2:]).to(
+                    torch.bool
+                )[0]
+                pos_l = self.position_embedding(
+                    shape=source.shape, device=device, dtype=pixel_values.dtype, mask=mask
+                ).to(source.dtype)
+                sources.append(source)
+                masks.append(mask)
+                position_embeddings_list.append(pos_l)
+
+        # Create queries
+        query_embeds = None
+        if not self.config.two_stage:
+            query_embeds = self.query_position_embeddings.weight
+
+        # Prepare encoder inputs (by flattening)
+        source_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes_list = []
+        for level, (source, mask, pos_embed) in enumerate(zip(sources, masks, position_embeddings_list)):
+            batch_size, num_channels, height, width = source.shape
+            spatial_shape = (height, width)
+            spatial_shapes_list.append(spatial_shape)
+            source = source.flatten(2).transpose(1, 2)
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)
+            mask = mask.flatten(1)
+            lvl_pos_embed = pos_embed + self.level_embed[level].view(1, 1, -1)
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            source_flatten.append(source)
+            mask_flatten.append(mask)
+        source_flatten = torch.cat(source_flatten, 1)
+        mask_flatten = torch.cat(mask_flatten, 1)
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
+        spatial_shapes = torch.as_tensor(spatial_shapes_list, dtype=torch.long, device=source_flatten.device)
+        level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
+        valid_ratios = torch.stack([self.get_valid_ratio(m, dtype=source_flatten.dtype) for m in masks], 1)
+
+        # Fourth, sent source_flatten + mask_flatten + lvl_pos_embed_flatten (backbone + proj layer output) through encoder
+        # Also provide spatial_shapes, level_start_index and valid_ratios
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                inputs_embeds=source_flatten,
+                attention_mask=mask_flatten,
+                spatial_position_embeddings=lvl_pos_embed_flatten,
+                spatial_shapes=spatial_shapes,
+                spatial_shapes_list=spatial_shapes_list,
+                level_start_index=level_start_index,
+                valid_ratios=valid_ratios,
+                **kwargs,
+            )
+
+        # Fifth, prepare decoder inputs
+        batch_size, _, num_channels = encoder_outputs[0].shape
+        enc_outputs_class = None
+        enc_outputs_coord_logits = None
+        if self.config.two_stage:
+            object_query_embedding, output_proposals = self.gen_encoder_output_proposals(
+                encoder_outputs[0], ~mask_flatten, spatial_shapes_list
+            )
+
+            # hack implementation for two-stage Deformable DETR
+            # apply a detection head to each pixel (A.4 in paper)
+            # linear projection for bounding box binary classification (i.e. foreground and background)
+            enc_outputs_class = self.decoder.class_embed[-1](object_query_embedding)
+            # 3-layer FFN to predict bounding boxes coordinates (bbox regression branch)
+            delta_bbox = self.decoder.bbox_embed[-1](object_query_embedding)
+            enc_outputs_coord_logits = delta_bbox + output_proposals
+
+            # only keep top scoring `config.two_stage_num_proposals` proposals
+            topk = self.config.two_stage_num_proposals
+            topk_proposals = torch.topk(enc_outputs_class[..., 0], topk, dim=1)[1]
+            topk_coords_logits = torch.gather(
+                enc_outputs_coord_logits,
+                1,
+                topk_proposals.unsqueeze(-1).repeat(1, 1, 4),
+            )
+
+            topk_coords_logits = topk_coords_logits.detach()
+            reference_points = topk_coords_logits.sigmoid()
+            init_reference_points = reference_points
+            pos_trans_out = self.pos_trans_norm(self.pos_trans(self.get_proposal_pos_embed(topk_coords_logits)))
+            query_embed, target = torch.split(pos_trans_out, num_channels, dim=2)
+        else:
+            query_embed, target = torch.split(query_embeds, num_channels, dim=1)
+            query_embed = query_embed.unsqueeze(0).expand(batch_size, -1, -1)
+            target = target.unsqueeze(0).expand(batch_size, -1, -1)
+            reference_points = self.reference_points(query_embed).sigmoid()
+            init_reference_points = reference_points
+
+        decoder_outputs = self.decoder(
+            inputs_embeds=target,
+            object_queries_position_embeddings=query_embed,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=mask_flatten,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
+            level_start_index=level_start_index,
+            valid_ratios=valid_ratios,
+            **kwargs,
+        )
+
+        return DeformableDetrModelOutput(
+            init_reference_points=init_reference_points,
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            intermediate_hidden_states=decoder_outputs.intermediate_hidden_states,
+            intermediate_reference_points=decoder_outputs.intermediate_reference_points,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+            enc_outputs_class=enc_outputs_class,
+            enc_outputs_coord_logits=enc_outputs_coord_logits,
+        )
+
+
+class DeformableDetrMLPPredictionHead(nn.Module):
+    """
+    Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates,
+    height and width of a bounding box w.r.t. an image.
+
+    """
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x
+
+
+@auto_docstring(
+    custom_intro="""
+    Deformable DETR Model (consisting of a backbone and encoder-decoder Transformer) with object detection heads on
+    top, for tasks such as COCO detection.
+    """
+)
+class DeformableDetrForObjectDetection(DeformableDetrPreTrainedModel):
+    # When using clones, all layers > 0 will be clones, but layer 0 *is* required
+    # We can't initialize the model on meta device as some weights are modified during the initialization
+    _no_split_modules = None
+    _tied_weights_keys = {
+        r"bbox_embed.(?![0])\d+": "bbox_embed.0",
+        r"class_embed.(?![0])\d+": "class_embed.0",
+    }
+
+    def __init__(self, config: DeformableDetrConfig):
+        super().__init__(config)
+        # Deformable DETR encoder-decoder model
+        self.model = DeformableDetrModel(config)
+        # Detection heads on top
+        # if two-stage, the last class_embed and bbox_embed is for region proposal generation
+        num_pred = (config.decoder_layers + 1) if config.two_stage else config.decoder_layers
+        self.class_embed = nn.ModuleList([nn.Linear(config.d_model, config.num_labels) for _ in range(num_pred)])
+        self.bbox_embed = nn.ModuleList(
+            [
+                DeformableDetrMLPPredictionHead(
+                    input_dim=config.d_model,
+                    hidden_dim=config.d_model,
+                    output_dim=4,
+                    num_layers=3,
+                )
+                for _ in range(num_pred)
+            ]
+        )
+        # Convert to instance attribute before modifying
+        self._tied_weights_keys = self._tied_weights_keys.copy()
+        if config.with_box_refine:
+            self.model.decoder.bbox_embed = self.bbox_embed
+            self._tied_weights_keys["bbox_embed"] = "model.decoder.bbox_embed"
+        if config.two_stage:
+            self.model.decoder.class_embed = self.class_embed
+            self._tied_weights_keys["class_embed"] = "model.decoder.class_embed"
+        self.post_init()
+
+    @auto_docstring
+    @can_return_tuple
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        pixel_mask: torch.LongTensor | None = None,
+        decoder_attention_mask: torch.FloatTensor | None = None,
+        encoder_outputs: torch.FloatTensor | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        decoder_inputs_embeds: torch.FloatTensor | None = None,
+        labels: list[dict] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.FloatTensor] | DeformableDetrObjectDetectionOutput:
+        r"""
+        decoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, num_queries)`, *optional*):
+            Not used by default. Can be used to mask object queries.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
+            can choose to directly pass a flattened representation of an image.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
+            embedded representation.
+        labels (`list[Dict]` of len `(batch_size,)`, *optional*):
+            Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the
+            following 2 keys: 'class_labels' and 'boxes' (the class labels and bounding boxes of an image in the batch
+            respectively). The class labels themselves should be a `torch.LongTensor` of len `(number of bounding boxes
+            in the image,)` and the boxes a `torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)`.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, DeformableDetrForObjectDetection
+        >>> from PIL import Image
+        >>> import httpx
+        >>> from io imoprt BytesIO
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("SenseTime/deformable-detr")
+        >>> model = DeformableDetrForObjectDetection.from_pretrained("SenseTime/deformable-detr")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> # convert outputs (bounding boxes and class logits) to Pascal VOC format (xmin, ymin, xmax, ymax)
+        >>> target_sizes = torch.tensor([image.size[::-1]])
+        >>> results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[
+        ...     0
+        ... ]
+        >>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
+        ...     box = [round(i, 2) for i in box.tolist()]
+        ...     print(
+        ...         f"Detected {model.config.id2label[label.item()]} with confidence "
+        ...         f"{round(score.item(), 3)} at location {box}"
+        ...     )
+        Detected cat with confidence 0.8 at location [16.5, 52.84, 318.25, 470.78]
+        Detected cat with confidence 0.789 at location [342.19, 24.3, 640.02, 372.25]
+        Detected remote with confidence 0.633 at location [40.79, 72.78, 176.76, 117.25]
+        ```"""
+        # First, sent images through DETR base model to obtain encoder + decoder outputs
+        outputs = self.model(
+            pixel_values,
+            pixel_mask=pixel_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            encoder_outputs=encoder_outputs,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            **kwargs,
+        )
+
+        hidden_states = outputs.intermediate_hidden_states
+        init_reference = outputs.init_reference_points
+        inter_references = outputs.intermediate_reference_points
+
+        # class logits + predicted bounding boxes
+        outputs_classes = []
+        outputs_coords = []
+
+        for level in range(hidden_states.shape[1]):
+            if level == 0:
+                reference = init_reference
+            else:
+                reference = inter_references[:, level - 1]
+            reference = inverse_sigmoid(reference)
+            outputs_class = self.class_embed[level](hidden_states[:, level])
+            delta_bbox = self.bbox_embed[level](hidden_states[:, level])
+            if reference.shape[-1] == 4:
+                outputs_coord_logits = delta_bbox + reference
+            elif reference.shape[-1] == 2:
+                delta_bbox[..., :2] += reference
+                outputs_coord_logits = delta_bbox
+            else:
+                raise ValueError(f"reference.shape[-1] should be 4 or 2, but got {reference.shape[-1]}")
+            outputs_coord = outputs_coord_logits.sigmoid()
+            outputs_classes.append(outputs_class)
+            outputs_coords.append(outputs_coord)
+        outputs_class = torch.stack(outputs_classes)
+        outputs_coord = torch.stack(outputs_coords)
+
+        logits = outputs_class[-1]
+        pred_boxes = outputs_coord[-1]
+
+        loss, loss_dict, auxiliary_outputs = None, None, None
+        if labels is not None:
+            loss, loss_dict, auxiliary_outputs = self.loss_function(
+                logits,
+                labels,
+                self.device,
+                pred_boxes,
+                self.config,
+                outputs_class,
+                outputs_coord,
+            )
+
+        return DeformableDetrObjectDetectionOutput(
+            loss=loss,
+            loss_dict=loss_dict,
+            logits=logits,
+            pred_boxes=pred_boxes,
+            auxiliary_outputs=auxiliary_outputs,
+            last_hidden_state=outputs.last_hidden_state,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+            intermediate_hidden_states=outputs.intermediate_hidden_states,
+            intermediate_reference_points=outputs.intermediate_reference_points,
+            init_reference_points=outputs.init_reference_points,
+            enc_outputs_class=outputs.enc_outputs_class,
+            enc_outputs_coord_logits=outputs.enc_outputs_coord_logits,
+        )
+
+
+__all__ = ["DeformableDetrForObjectDetection", "DeformableDetrModel", "DeformableDetrPreTrainedModel"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/dpt/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_dpt import *
+    from .image_processing_dpt import *
+    from .image_processing_pil_dpt import *
+    from .modeling_dpt import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/dpt/configuration_dpt.py

@@ -0,0 +1,158 @@
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""DPT model configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...backbone_utils import consolidate_backbone_kwargs_to_config
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+from ..auto.configuration_auto import AutoConfig
+
+
+@auto_docstring(checkpoint="Intel/dpt-large")
+@strict
+class DPTConfig(PreTrainedConfig):
+    r"""
+    is_hybrid (`bool`, *optional*, defaults to `False`):
+        Whether to use a hybrid backbone. Useful in the context of loading DPT-Hybrid models.
+    backbone_out_indices (`list[int]`, *optional*, defaults to `[2, 5, 8, 11]`):
+        Indices of the intermediate hidden states to use from backbone.
+    readout_type (`str`, *optional*, defaults to `"project"`):
+        The readout type to use when processing the readout token (CLS token) of the intermediate hidden states of
+        the ViT backbone. Can be one of [`"ignore"`, `"add"`, `"project"`].
+        - "ignore" simply ignores the CLS token.
+        - "add" passes the information from the CLS token to all other tokens by adding the representations.
+        - "project" passes information to the other tokens by concatenating the readout to all other tokens before
+          projecting the
+        representation to the original feature dimension D using a linear layer followed by a GELU non-linearity.
+    reassemble_factors (`list[int]`, *optional*, defaults to `[4, 2, 1, 0.5]`):
+        The up/downsampling factors of the reassemble layers.
+    neck_hidden_sizes (`list[str]`, *optional*, defaults to `[96, 192, 384, 768]`):
+        The hidden sizes to project to for the feature maps of the backbone.
+    fusion_hidden_size (`int`, *optional*, defaults to 256):
+        The number of channels before fusion.
+    head_in_index (`int`, *optional*, defaults to -1):
+        The index of the features to use in the heads.
+    use_batch_norm_in_fusion_residual (`bool`, *optional*, defaults to `False`):
+        Whether to use batch normalization in the pre-activate residual units of the fusion blocks.
+    use_bias_in_fusion_residual (`bool`, *optional*, defaults to `True`):
+        Whether to use bias in the pre-activate residual units of the fusion blocks.
+    add_projection (`bool`, *optional*, defaults to `False`):
+        Whether to add a projection layer before the depth estimation head.
+    use_auxiliary_head (`bool`, *optional*, defaults to `True`):
+        Whether to use an auxiliary head during training.
+    auxiliary_loss_weight (`float`, *optional*, defaults to 0.4):
+        Weight of the cross-entropy loss of the auxiliary head.
+    semantic_classifier_dropout (`float`, *optional*, defaults to 0.1):
+        The dropout ratio for the semantic classification head.
+    backbone_featmap_shape (`list[int]`, *optional*, defaults to `[1, 1024, 24, 24]`):
+        Used only for the `hybrid` embedding type. The shape of the feature maps of the backbone.
+    neck_ignore_stages (`list[int]`, *optional*, defaults to `[0, 1]`):
+        Used only for the `hybrid` embedding type. The stages of the readout layers to ignore.
+    pooler_output_size (`int`, *optional*):
+        Dimensionality of the pooler layer. If None, defaults to `hidden_size`.
+    pooler_act (`str`, *optional*, defaults to `"tanh"`):
+        The activation function to be used by the pooler.
+
+    Example:
+
+    ```python
+    >>> from transformers import DPTModel, DPTConfig
+
+    >>> # Initializing a DPT dpt-large style configuration
+    >>> configuration = DPTConfig()
+
+    >>> # Initializing a model from the dpt-large style configuration
+    >>> model = DPTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "dpt"
+    sub_configs = {"backbone_config": AutoConfig}
+
+    # NOTE: some values are typed as `None` on purpose
+    # DPT creates one of: backbone or the general model only
+    # so official checkpoint saved them as `None`
+    hidden_size: int = 768
+    num_hidden_layers: None | int = 12
+    num_attention_heads: int | None = 12
+    intermediate_size: int | None = 3072
+    hidden_act: str = "gelu"
+    hidden_dropout_prob: float | int | None = 0.0
+    attention_probs_dropout_prob: float | int | None = 0.0
+    initializer_range: float = 0.02
+    layer_norm_eps: float | None = 1e-12
+    image_size: int | list[int] | tuple[int, int] | None = 384
+    patch_size: int | list[int] | tuple[int, int] | None = 16
+    num_channels: int | None = 3
+    is_hybrid: bool = False
+    qkv_bias: bool | None = True
+    backbone_out_indices: list[int] | tuple[int, ...] | None = (2, 5, 8, 11)
+    readout_type: str = "project"
+    reassemble_factors: list[int | float] | tuple[int | float, ...] = (4, 2, 1, 0.5)
+    neck_hidden_sizes: list[int] | tuple[int, ...] = (96, 192, 384, 768)
+    fusion_hidden_size: int = 256
+    head_in_index: int = -1
+    use_batch_norm_in_fusion_residual: bool | None = False
+    use_bias_in_fusion_residual: bool | None = None
+    add_projection: bool = False
+    use_auxiliary_head: bool | None = True
+    auxiliary_loss_weight: float = 0.4
+    semantic_loss_ignore_index: int = 255
+    semantic_classifier_dropout: float | int = 0.1
+    backbone_featmap_shape: list[int] | tuple[int, ...] | None = (1, 1024, 24, 24)
+    neck_ignore_stages: list[int] | tuple[int, ...] = (0, 1)
+    backbone_config: dict | PreTrainedConfig | None = None
+    pooler_output_size: int | None = None
+    pooler_act: str = "tanh"
+
+    def __post_init__(self, **kwargs):
+        if self.readout_type not in ["ignore", "add", "project"]:
+            raise ValueError("Readout_type must be one of ['ignore', 'add', 'project']")
+
+        if self.is_hybrid:
+            if isinstance(self.backbone_config, dict):
+                self.backbone_config.setdefault("model_type", "bit")
+
+            self.backbone_config, kwargs = consolidate_backbone_kwargs_to_config(
+                backbone_config=self.backbone_config,
+                default_config_type="bit",
+                default_config_kwargs={
+                    "global_padding": "same",
+                    "layer_type": "bottleneck",
+                    "depths": [3, 4, 9],
+                    "out_features": ["stage1", "stage2", "stage3"],
+                    "embedding_dynamic_padding": True,
+                },
+                **kwargs,
+            )
+            if self.readout_type != "project":
+                raise ValueError("Readout type must be 'project' when using `DPT-hybrid` mode.")
+        elif kwargs.get("backbone") is not None or self.backbone_config is not None:
+            self.backbone_config, kwargs = consolidate_backbone_kwargs_to_config(
+                backbone_config=self.backbone_config,
+                **kwargs,
+            )
+            self.backbone_out_indices = None
+
+        self.backbone_featmap_shape = self.backbone_featmap_shape if self.is_hybrid else None
+        self.neck_ignore_stages = self.neck_ignore_stages if self.is_hybrid else []
+        self.pooler_output_size = self.pooler_output_size if self.pooler_output_size else self.hidden_size
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["DPTConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/dpt/image_processing_pil_dpt.py

@@ -0,0 +1,312 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for DPT."""
+
+import math
+from collections.abc import Iterable
+from typing import TYPE_CHECKING
+
+import numpy as np
+
+from ...image_processing_backends import PilBackend
+from ...image_processing_utils import BatchFeature
+from ...image_transforms import pad as np_pad
+from ...image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+)
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring, logging
+from ...utils.import_utils import requires
+
+
+if TYPE_CHECKING:
+    from ...modeling_outputs import DepthEstimatorOutput
+
+logger = logging.get_logger(__name__)
+
+
+# Adapted from transformers.models.dpt.image_processing_dpt.DPTImageProcessorKwargs
+class DPTImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    ensure_multiple_of (`int`, *optional*, defaults to 1):
+        If `do_resize` is `True`, the image is resized to a size that is a multiple of this value. Can be overridden
+        by `ensure_multiple_of` in `preprocess`.
+    keep_aspect_ratio (`bool`, *optional*, defaults to `False`):
+        If `True`, the image is resized to the largest possible size such that the aspect ratio is preserved. Can
+        be overridden by `keep_aspect_ratio` in `preprocess`.
+    do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`):
+        Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0
+        is used for background, and background itself is not included in all classes of a dataset (e.g.
+        ADE20k). The background label will be replaced by 255.
+    """
+
+    ensure_multiple_of: int
+    size_divisor: int
+    keep_aspect_ratio: bool
+    do_reduce_labels: bool
+
+
+# Adapted from transformers.models.dpt.image_processing_dpt.get_resize_output_image_size
+def get_resize_output_image_size(
+    input_image: np.ndarray,
+    output_size: int | Iterable[int],
+    keep_aspect_ratio: bool,
+    multiple: int,
+) -> SizeDict:
+    def constrain_to_multiple_of(val, multiple, min_val=0, max_val=None):
+        x = round(val / multiple) * multiple
+
+        if max_val is not None and x > max_val:
+            x = math.floor(val / multiple) * multiple
+
+        if x < min_val:
+            x = math.ceil(val / multiple) * multiple
+
+        return x
+
+    input_height, input_width = input_image.shape[-2:]
+    output_height, output_width = output_size
+
+    # determine new height and width
+    scale_height = output_height / input_height
+    scale_width = output_width / input_width
+
+    if keep_aspect_ratio:
+        # scale as little as possible
+        if abs(1 - scale_width) < abs(1 - scale_height):
+            # fit width
+            scale_height = scale_width
+        else:
+            # fit height
+            scale_width = scale_height
+
+    new_height = constrain_to_multiple_of(scale_height * input_height, multiple=multiple)
+    new_width = constrain_to_multiple_of(scale_width * input_width, multiple=multiple)
+
+    return SizeDict(height=new_height, width=new_width)
+
+
+@auto_docstring
+class DPTImageProcessorPil(PilBackend):
+    """PIL backend for DPT with custom resize and pad."""
+
+    valid_kwargs = DPTImageProcessorKwargs
+
+    resample = PILImageResampling.BICUBIC
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    size = {"height": 384, "width": 384}
+    default_to_square = True
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_pad = False
+    rescale_factor = 1 / 255
+    ensure_multiple_of = 1
+    keep_aspect_ratio = False
+
+    def __init__(self, **kwargs: Unpack[DPTImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @auto_docstring
+    def preprocess(
+        self,
+        images: ImageInput,
+        segmentation_maps: ImageInput | None = None,
+        **kwargs: Unpack[DPTImageProcessorKwargs],
+    ) -> BatchFeature:
+        r"""
+        segmentation_maps (`ImageInput`, *optional*):
+            The segmentation maps to preprocess.
+        """
+        return super().preprocess(images, segmentation_maps, **kwargs)
+
+    def _preprocess_image_like_inputs(
+        self,
+        images: ImageInput,
+        segmentation_maps: ImageInput | None,
+        do_convert_rgb: bool,
+        input_data_format: ChannelDimension,
+        return_tensors: str | TensorType | None,
+        device: str | None = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """Handle extra inputs beyond images."""
+        images = self._prepare_image_like_inputs(
+            images=images, do_convert_rgb=do_convert_rgb, input_data_format=input_data_format, device=device
+        )
+        images_kwargs = kwargs.copy()
+        images_kwargs["do_reduce_labels"] = False
+        data = {}
+        data["pixel_values"] = self._preprocess(images, **images_kwargs)
+
+        if segmentation_maps is not None:
+            processed_segmentation_maps = self._prepare_image_like_inputs(
+                images=segmentation_maps,
+                expected_ndims=2,
+                do_convert_rgb=False,
+                input_data_format=ChannelDimension.FIRST,
+            )
+
+            segmentation_maps_kwargs = kwargs.copy()
+            segmentation_maps_kwargs.update({"do_normalize": False, "do_rescale": False})
+            processed_segmentation_maps = self._preprocess(
+                images=processed_segmentation_maps, **segmentation_maps_kwargs
+            )
+
+            processed_segmentation_maps = [
+                processed_segmentation_map.squeeze(0).astype(np.int64)
+                for processed_segmentation_map in processed_segmentation_maps
+            ]
+            data["labels"] = processed_segmentation_maps
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    def reduce_label(self, image: np.ndarray) -> np.ndarray:
+        """Reduce label values by 1, replacing 0 with 255."""
+        image[image == 0] = 255
+        image = image - 1
+        image[image == 254] = 255
+        return image
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: SizeDict,
+        resample: PILImageResampling | None,
+        ensure_multiple_of: int = 1,
+        keep_aspect_ratio: bool = False,
+        **kwargs,
+    ) -> np.ndarray:
+        """Resize with aspect ratio and multiple constraints."""
+        if size.height is None or size.width is None:
+            raise ValueError(f"Size must contain 'height' and 'width' keys. Got {size.keys()}")
+        output_size = get_resize_output_image_size(
+            image,
+            output_size=(size.height, size.width),
+            keep_aspect_ratio=keep_aspect_ratio,
+            multiple=ensure_multiple_of,
+        )
+        return super().resize(image, output_size, resample, **kwargs)
+
+    def pad_image(self, image: np.ndarray, size_divisor: int = 1, **kwargs) -> np.ndarray:
+        """Center pad image to be a multiple of size_divisor."""
+
+        def _get_pad(size, size_divisor):
+            new_size = math.ceil(size / size_divisor) * size_divisor
+            pad_size = new_size - size
+            pad_size_left = pad_size // 2
+            pad_size_right = pad_size - pad_size_left
+            return pad_size_left, pad_size_right
+
+        height, width = image.shape[-2:]
+        pad_size_left, pad_size_right = _get_pad(height, size_divisor)
+        pad_size_top, pad_size_bottom = _get_pad(width, size_divisor)
+        return np_pad(
+            image,
+            padding=((pad_size_left, pad_size_right), (pad_size_top, pad_size_bottom)),
+            data_format=ChannelDimension.FIRST,
+            input_data_format=ChannelDimension.FIRST,
+        )
+
+    def _preprocess(
+        self,
+        images: list[np.ndarray],
+        do_resize: bool,
+        size: SizeDict,
+        resample: PILImageResampling | None,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: float | list[float] | None,
+        image_std: float | list[float] | None,
+        do_pad: bool | None,
+        do_reduce_labels: bool,
+        keep_aspect_ratio: bool,
+        ensure_multiple_of: int,
+        size_divisor: int | None = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """Custom preprocessing for DPT."""
+        processed_images = []
+        for image in images:
+            if do_reduce_labels:
+                image = self.reduce_label(image)
+            if do_resize:
+                image = self.resize(
+                    image, size, resample, ensure_multiple_of=ensure_multiple_of, keep_aspect_ratio=keep_aspect_ratio
+                )
+            if do_rescale:
+                image = self.rescale(image, rescale_factor)
+            if do_normalize:
+                image = self.normalize(image, image_mean, image_std)
+            if do_pad and size_divisor is not None:
+                image = self.pad_image(image, size_divisor)
+            processed_images.append(image)
+        return processed_images
+
+    @requires(backends=("torch",))
+    def post_process_semantic_segmentation(self, outputs, target_sizes: list[tuple] | None = None):
+        """Converts the output of [`DPTForSemanticSegmentation`] into semantic segmentation maps."""
+        import torch
+
+        logits = outputs.logits
+        if target_sizes is not None:
+            if len(logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+            if isinstance(target_sizes, torch.Tensor):
+                target_sizes = target_sizes.numpy()
+            semantic_segmentation = []
+            for idx in range(len(logits)):
+                resized_logits = torch.nn.functional.interpolate(
+                    logits[idx].unsqueeze(dim=0), size=target_sizes[idx], mode="bilinear", align_corners=False
+                )
+                semantic_map = resized_logits[0].argmax(dim=0)
+                semantic_segmentation.append(semantic_map)
+        else:
+            semantic_segmentation = logits.argmax(dim=1)
+            semantic_segmentation = [semantic_segmentation[i] for i in range(semantic_segmentation.shape[0])]
+        return semantic_segmentation
+
+    @requires(backends=("torch",))
+    def post_process_depth_estimation(
+        self, outputs: "DepthEstimatorOutput", target_sizes: TensorType | list[tuple[int, int]] | None = None
+    ) -> list[dict[str, TensorType]]:
+        """Converts the raw output of [`DepthEstimatorOutput`] into final depth predictions."""
+        import torch
+
+        predicted_depth = outputs.predicted_depth
+        if (target_sizes is not None) and (len(predicted_depth) != len(target_sizes)):
+            raise ValueError(
+                "Make sure that you pass in as many target sizes as the batch dimension of the predicted depth"
+            )
+        results = []
+        target_sizes = [None] * len(predicted_depth) if target_sizes is None else target_sizes
+        for depth, target_size in zip(predicted_depth, target_sizes):
+            if target_size is not None:
+                depth = torch.nn.functional.interpolate(
+                    depth.unsqueeze(0).unsqueeze(1), size=target_size, mode="bicubic", align_corners=False
+                ).squeeze()
+            results.append({"predicted_depth": depth})
+        return results
+
+
+__all__ = ["DPTImageProcessorPil"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/__init__.py

@@ -0,0 +1,339 @@
+#!/usr/bin/env python
+
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import lru_cache
+
+from packaging import version
+
+from .. import __version__
+from .auto_docstring import (
+    ClassAttrs,
+    ClassDocstring,
+    ImageProcessorArgs,
+    ModelArgs,
+    ModelOutputArgs,
+    auto_class_docstring,
+    auto_docstring,
+    get_args_doc_from_source,
+    parse_docstring,
+    set_min_indent,
+)
+from .chat_template_utils import DocstringParsingException, TypeHintParsingException, get_json_schema
+from .constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD
+from .doc import (
+    add_code_sample_docstrings,
+    add_end_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    copy_func,
+    replace_return_docstrings,
+)
+from .generic import (
+    ContextManagers,
+    ExplicitEnum,
+    ModelOutput,
+    PaddingStrategy,
+    TensorType,
+    TransformersKwargs,
+    _is_tensor_or_array_like,
+    can_return_loss,
+    can_return_tuple,
+    expand_dims,
+    filter_out_non_signature_kwargs,
+    find_labels,
+    flatten_dict,
+    is_numpy_array,
+    is_tensor,
+    is_timm_config_dict,
+    is_timm_local_checkpoint,
+    is_torch_device,
+    is_torch_dtype,
+    is_torch_tensor,
+    reshape,
+    safe_load_json_file,
+    squeeze,
+    strtobool,
+    tensor_size,
+    to_numpy,
+    to_py_obj,
+    torch_float,
+    torch_int,
+    transpose,
+)
+from .hub import (
+    CHAT_TEMPLATE_DIR,
+    CHAT_TEMPLATE_FILE,
+    CLOUDFRONT_DISTRIB_PREFIX,
+    HF_MODULES_CACHE,
+    LEGACY_PROCESSOR_CHAT_TEMPLATE_FILE,
+    S3_BUCKET_PREFIX,
+    TRANSFORMERS_DYNAMIC_MODULE_NAME,
+    EntryNotFoundError,
+    PushInProgress,
+    PushToHubMixin,
+    RepositoryNotFoundError,
+    RevisionNotFoundError,
+    cached_file,
+    define_sagemaker_information,
+    extract_commit_hash,
+    has_file,
+    http_user_agent,
+    list_repo_templates,
+    try_to_load_from_cache,
+)
+from .import_utils import (
+    ACCELERATE_MIN_VERSION,
+    BITSANDBYTES_MIN_VERSION,
+    ENV_VARS_TRUE_AND_AUTO_VALUES,
+    ENV_VARS_TRUE_VALUES,
+    GGUF_MIN_VERSION,
+    TRITON_MIN_VERSION,
+    XLA_FSDPV2_MIN_VERSION,
+    DummyObject,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    check_torch_load_is_safe,
+    direct_transformers_import,
+    enable_tf32,
+    get_torch_version,
+    is_accelerate_available,
+    is_apex_available,
+    is_apollo_torch_available,
+    is_aqlm_available,
+    is_auto_round_available,
+    is_av_available,
+    is_bitsandbytes_available,
+    is_bs4_available,
+    is_causal_conv1d_available,
+    is_coloredlogs_available,
+    is_compressed_tensors_available,
+    is_cuda_platform,
+    is_cv2_available,
+    is_cython_available,
+    is_datasets_available,
+    is_decord_available,
+    is_detectron2_available,
+    is_env_variable_false,
+    is_env_variable_true,
+    is_essentia_available,
+    is_faiss_available,
+    is_fbgemm_gpu_available,
+    is_flash_attn_2_available,
+    is_flash_attn_3_available,
+    is_flash_attn_4_available,
+    is_flash_attn_greater_or_equal,
+    is_flash_attn_greater_or_equal_2_10,
+    is_flash_linear_attention_available,
+    is_flute_available,
+    is_fouroversix_available,
+    is_fp_quant_available,
+    is_fsdp_available,
+    is_g2p_en_available,
+    is_galore_torch_available,
+    is_gguf_available,
+    is_gptqmodel_available,
+    is_grokadamw_available,
+    is_grouped_mm_available,
+    is_habana_gaudi1,
+    is_hadamard_available,
+    is_hqq_available,
+    is_huggingface_hub_greater_or_equal,
+    is_in_notebook,
+    is_ipython_available,
+    is_jinja_available,
+    is_jmespath_available,
+    is_jumanpp_available,
+    is_kenlm_available,
+    is_kernels_available,
+    is_levenshtein_available,
+    is_libcst_available,
+    is_librosa_available,
+    is_liger_kernel_available,
+    is_llm_awq_available,
+    is_lomo_available,
+    is_matplotlib_available,
+    is_mistral_common_available,
+    is_mlx_available,
+    is_multipart_available,
+    is_natten_available,
+    is_ninja_available,
+    is_nltk_available,
+    is_num2words_available,
+    is_numba_available,
+    is_onnx_available,
+    is_openai_available,
+    is_optimum_available,
+    is_optimum_quanto_available,
+    is_pandas_available,
+    is_peft_available,
+    is_phonemizer_available,
+    is_pretty_midi_available,
+    is_protobuf_available,
+    is_psutil_available,
+    is_py3nvml_available,
+    is_pyctcdecode_available,
+    is_pytesseract_available,
+    is_pytest_available,
+    is_pytest_order_available,
+    is_pytorch_quantization_available,
+    is_quanto_greater,
+    is_quark_available,
+    is_qutlass_available,
+    is_rich_available,
+    is_rjieba_available,
+    is_rocm_platform,
+    is_sacremoses_available,
+    is_sagemaker_dp_enabled,
+    is_sagemaker_mp_enabled,
+    is_schedulefree_available,
+    is_scipy_available,
+    is_sentencepiece_available,
+    is_seqio_available,
+    is_serve_available,
+    is_sinq_available,
+    is_sklearn_available,
+    is_soundfile_available,
+    is_spacy_available,
+    is_speech_available,
+    is_spqr_available,
+    is_sudachi_available,
+    is_sudachi_projection_available,
+    is_tiktoken_available,
+    is_timm_available,
+    is_tokenizers_available,
+    is_torch_accelerator_available,
+    is_torch_available,
+    is_torch_bf16_available_on_device,
+    is_torch_bf16_gpu_available,
+    is_torch_cuda_available,
+    is_torch_deterministic,
+    is_torch_flex_attn_available,
+    is_torch_fp16_available_on_device,
+    is_torch_fx_proxy,
+    is_torch_greater_or_equal,
+    is_torch_hpu_available,
+    is_torch_mlu_available,
+    is_torch_mps_available,
+    is_torch_musa_available,
+    is_torch_neuron_available,
+    is_torch_neuroncore_available,
+    is_torch_npu_available,
+    is_torch_optimi_available,
+    is_torch_tensorrt_fx_available,
+    is_torch_tf32_available,
+    is_torch_tpu_available,
+    is_torch_xla_available,
+    is_torch_xpu_available,
+    is_torchao_available,
+    is_torchaudio_available,
+    is_torchcodec_available,
+    is_torchdistx_available,
+    is_torchdynamo_compiling,
+    is_torchdynamo_exporting,
+    is_torchvision_available,
+    is_torchvision_v2_available,
+    is_tracing,
+    is_training_run_on_sagemaker,
+    is_triton_available,
+    is_uroman_available,
+    is_vision_available,
+    is_vptq_available,
+    is_xlstm_available,
+    is_yt_dlp_available,
+    requires_backends,
+    torch_compilable_check,
+    torch_only_method,
+)
+from .kernel_config import KernelConfig
+from .peft_utils import (
+    ADAPTER_CONFIG_NAME,
+    ADAPTER_SAFE_WEIGHTS_NAME,
+    ADAPTER_WEIGHTS_NAME,
+    check_peft_version,
+    find_adapter_config_file,
+)
+
+
+WEIGHTS_NAME = "pytorch_model.bin"
+WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
+SAFE_WEIGHTS_NAME = "model.safetensors"
+SAFE_WEIGHTS_INDEX_NAME = "model.safetensors.index.json"
+CONFIG_NAME = "config.json"
+FEATURE_EXTRACTOR_NAME = "preprocessor_config.json"
+IMAGE_PROCESSOR_NAME = "preprocessor_config.json"
+VIDEO_PROCESSOR_NAME = "video_preprocessor_config.json"
+AUDIO_TOKENIZER_NAME = "audio_tokenizer_config.json"
+PROCESSOR_NAME = "processor_config.json"
+GENERATION_CONFIG_NAME = "generation_config.json"
+MODEL_CARD_NAME = "modelcard.json"
+
+
+SENTENCEPIECE_UNDERLINE = "▁"
+SPIECE_UNDERLINE = SENTENCEPIECE_UNDERLINE  # Kept for backward compatibility
+
+MULTIPLE_CHOICE_DUMMY_INPUTS = [
+    [[0, 1, 0, 1], [1, 0, 0, 1]]
+] * 2  # Needs to have 0s and 1s only since XLM uses it for langs too.
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
+
+
+def check_min_version(min_version):
+    if version.parse(__version__) < version.parse(min_version):
+        if "dev" in min_version:
+            error_message = (
+                "This example requires a source install from HuggingFace Transformers (see "
+                "`https://huggingface.co/docs/transformers/installation#install-from-source`),"
+            )
+        else:
+            error_message = f"This example requires a minimum version of {min_version},"
+        error_message += f" but the version found is {__version__}.\n"
+        raise ImportError(
+            error_message
+            + "Check out https://github.com/huggingface/transformers/tree/main/examples#important-note for the examples corresponding to other "
+            "versions of HuggingFace Transformers."
+        )
+
+
+@lru_cache
+def get_available_devices() -> frozenset[str]:
+    """
+    Returns a frozenset of devices available for the current PyTorch installation.
+    """
+    devices = {"cpu"}  # `cpu` is always supported as a device in PyTorch
+
+    if is_torch_cuda_available():
+        devices.add("cuda")
+
+    if is_torch_mps_available():
+        devices.add("mps")
+
+    if is_torch_xpu_available():
+        devices.add("xpu")
+
+    if is_torch_npu_available():
+        devices.add("npu")
+
+    if is_torch_hpu_available():
+        devices.add("hpu")
+
+    if is_torch_mlu_available():
+        devices.add("mlu")
+
+    if is_torch_musa_available():
+        devices.add("musa")
+
+    return frozenset(devices)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/backbone_utils.py

@@ -0,0 +1,19 @@
+import warnings
+
+from ..backbone_utils import BackboneConfigMixin, BackboneMixin
+
+
+class BackboneConfigMixin(BackboneConfigMixin):
+    warnings.warn(
+        "Importing `BackboneConfigMixin` from `utils/backbone_utils.py` is deprecated and will be removed in "
+        "Transformers v5.10. Import as `from transformers.backbone_utils import BackboneConfigMixin` instead.",
+        FutureWarning,
+    )
+
+
+class BackboneMixin(BackboneMixin):
+    warnings.warn(
+        "Importing `BackboneMixin` from `utils/backbone_utils.py` is deprecated and will be removed in "
+        "Transformers v5.10. Import as `from transformers.backbone_utils import BackboneMixin` instead.",
+        FutureWarning,
+    )

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/chat_parsing_utils.py

@@ -0,0 +1,305 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import json
+import re
+
+from transformers.utils import is_jmespath_available
+
+
+if is_jmespath_available():
+    import jmespath
+else:
+    jmespath = None
+
+
+def _gemma4_json_to_json(text: str) -> str:
+    """Convert Gemma4 tool call format (unquoted keys, ``<|"|>`` string delimiters) to valid JSON."""
+    strings = []
+
+    def _capture(m):
+        strings.append(m.group(1))
+        return f"\x00{len(strings) - 1}\x00"
+
+    # Grab the inside of gemma-quotes and store them for later
+    text = re.sub(r'<\|"\|>(.*?)<\|"\|>', _capture, text, flags=re.DOTALL)
+    # Add quotes to the bare keys elsewhere
+    text = re.sub(r"(?<=[{,])(\w+):", r'"\1":', text)
+
+    # Put the inside of the quotes back afterwards
+    for i, s in enumerate(strings):
+        text = text.replace(f"\x00{i}\x00", json.dumps(s))
+
+    return text
+
+
+def _parse_re_match(node_match: re.Match) -> dict | str:
+    # If the regex has named groups, return a dict of those groups
+    if node_match.groupdict():
+        return {key: val for key, val in node_match.groupdict().items() if val is not None}
+    # Otherwise the regex must have exactly one unnamed group, and we return that
+    else:
+        groups = list(node_match.groups())
+        if len(groups) > 1:
+            raise ValueError(f"Regex has multiple unnamed groups!\nGroups: {groups}\n")
+        elif len(groups) == 0:
+            raise ValueError(f"Regex has no capture groups:\n\n{node_match.group(0)}")
+        return groups[0]
+
+
+def recursive_parse(
+    node_content: str | list | dict,
+    node_schema: dict,
+):
+    """
+    This function takes content and a JSON schema which includes
+    regex extractors, and recursively parses the content. The output
+    should be a data structure matching the schema.
+
+    Args:
+        node_content: The content corresponding to this node. Usually a string, but can be something else
+                      if the parent node has multiple capture groups or named groups. In that case,
+                      we generally pass the capture groups straight through to the children of this node
+                      and don't do any parsing at this level.
+        node_schema: The schema node controlling the parsing.
+
+    Returns:
+        The parsed data structure for the current node.
+    """
+
+    # If the schema has a const, we just return that value and do absolutely nothing else
+    if "const" in node_schema:
+        return node_schema["const"]
+
+    # If the node content is None, we return None. EZ.
+    if node_content is None:
+        return None
+
+    # If not, we have to do a little parsing. First, set some vars and do basic validation
+    node_type = node_schema.get("type")
+    has_regex = (
+        "x-regex" in node_schema
+        or "x-regex-iterator" in node_schema
+        or "x-regex-key-value" in node_schema
+        or "x-regex-substitutions" in node_schema
+    )
+    if has_regex and not isinstance(node_content, str):
+        raise TypeError(
+            "Schema node got a non-string input, but has a regex for parsing or substitution.\n"
+            f"Input: {node_content}\n"
+            f"Schema: {node_schema}"
+        )
+
+    node_subs = node_schema.get("x-regex-substitutions", [])
+    for node_sub in node_subs:
+        node_content = re.sub(node_sub[0], node_sub[1], node_content, flags=re.DOTALL)
+    node_regex = node_schema.get("x-regex")
+    node_regex_iterator = node_schema.get("x-regex-iterator")
+    node_regex_to_dict = node_schema.get("x-regex-key-value")
+    if node_regex is not None:
+        node_match = re.search(node_regex, node_content, flags=re.DOTALL)
+        if not node_match:
+            return None
+        node_content = _parse_re_match(node_match)
+    if node_regex_iterator is not None:
+        if node_type != "array":
+            raise TypeError(f"Schema node with type {node_type} cannot use x-regex-iterator.\nSchema: {node_schema}")
+        # Note that this can be applied after a standard node-regex search
+        node_content = [
+            _parse_re_match(node_match)
+            for node_match in re.finditer(node_regex_iterator, node_content, flags=re.DOTALL)
+        ]
+        if not node_content:
+            return None
+    if node_regex_to_dict is not None:
+        if node_type != "object":
+            raise TypeError(f"Schema node with type {node_type} cannot use x-regex-key-value.\nSchema: {node_schema}")
+        # Note that this can be applied after a standard node-regex search
+        output_content = {}
+        for node_match in re.finditer(node_regex_to_dict, node_content, flags=re.DOTALL):
+            match_groups = _parse_re_match(node_match)
+            if not isinstance(match_groups, dict) or "key" not in match_groups or "value" not in match_groups:
+                raise ValueError(
+                    f"Regex for x-regex-key-value must have named groups 'key' and 'value'.\n"
+                    f"Match groups: {match_groups}\n"
+                    f"Schema: {node_schema}"
+                )
+            output_content[match_groups["key"]] = match_groups["value"]
+        node_content = output_content
+        if not node_content:
+            return None
+
+    # Next, if the node has a parser, apply it. We do this after regexes so that the regex can extract
+    # a substring to parse, if needed.
+    if "x-parser" in node_schema:
+        parser = node_schema["x-parser"]
+        if parser == "gemma4-tool-call":
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Node has Gemma4 tool call parser but got non-string input: {node_content}\nSchema: {node_schema}"
+                )
+            node_content = _gemma4_json_to_json(node_content)
+            parser = "json"  # fall through to the JSON parser below - don't add an elif!
+        if parser == "json":
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Node has JSON parser but got non-string input: {node_content}\nSchema: {node_schema}"
+                )
+            parser_args = node_schema.get("x-parser-args", {})
+            transform = parser_args.get("transform")
+            allow_non_json = parser_args.get("allow_non_json", False)
+            try:
+                parsed_json = json.loads(node_content)
+            except json.JSONDecodeError as e:
+                if allow_non_json:
+                    parsed_json = node_content
+                else:
+                    raise ValueError(
+                        f"Node has JSON parser but could not parse its contents as JSON. You can use the `allow_non_json` parser arg for nodes which may contain JSON or string content.\n\nContent: {node_content}\n\nError: {e}"
+                    )
+            if transform is not None:
+                if jmespath is None:
+                    raise ImportError(
+                        "Chat response schema includes a jmespath transformation, but jmespath is not installed. You can install it with `pip install jmespath`."
+                    )
+                parsed_json = jmespath.search(parser_args["transform"], parsed_json)
+            node_content = parsed_json
+        else:
+            raise ValueError(f"Unknown parser {parser} for schema node: {node_schema}")
+
+    # Finally, handle parsed content based on schema type and recurse if required
+    if node_type == "object":
+        parsed_schema = {}
+        if isinstance(node_content, str):
+            # This means we don't have a regex at this level, so all of our child nodes need to parse the whole
+            # string themselves to extract their value.
+            if "properties" not in node_schema:
+                raise ValueError(
+                    f"Object node received string content but has no regex or parser to handle it.\n"
+                    f"Content: {node_content}\n"
+                    f"Schema: {node_schema}"
+                )
+            for key, child_node in node_schema["properties"].items():
+                child_node_content = recursive_parse(node_content, node_schema["properties"][key])
+                if child_node_content is not None:
+                    parsed_schema[key] = child_node_content
+        elif isinstance(node_content, dict):
+            for key, child_node in node_schema.get("properties", {}).items():
+                if "const" in child_node:
+                    parsed_schema[key] = child_node["const"]
+                elif key in node_content:
+                    parsed_schema[key] = recursive_parse(node_content[key], child_node)
+                elif "default" in child_node:
+                    parsed_schema[key] = child_node["default"]
+            additional_schema = node_schema.get("additionalProperties", True)
+            # We want to check only for False values; {} is "falsy" but should pass through
+            if additional_schema is not False:
+                additional_schema = additional_schema if isinstance(additional_schema, dict) else {}
+                for key, value in node_content.items():
+                    if key not in node_schema.get("properties", {}):
+                        parsed_schema[key] = recursive_parse(value, additional_schema)
+        else:
+            raise TypeError(f"Expected a dict or str for schema node with type object, got {node_content}")
+        required = node_schema.get("required", [])
+        missing = [key for key in required if key not in parsed_schema]
+        if missing:
+            input_preview = repr(node_content[:500]) if isinstance(node_content, str) else repr(node_content)
+            raise ValueError(
+                f"Required fields {missing} are missing from parsed output.\n"
+                f"Parsed: {parsed_schema}\n"
+                f"Input: {input_preview}"
+            )
+        return parsed_schema
+    elif node_type == "array":
+        if not node_content:
+            return []
+        parsed_schema = []
+        if "items" in node_schema:
+            if not isinstance(node_content, list):
+                raise TypeError(f"Expected a list or regex for schema node with type array, got {node_content}")
+            for item in node_content:
+                parsed_schema.append(recursive_parse(item, node_schema["items"]))
+            return parsed_schema
+        elif "prefixItems" in node_schema:
+            if not isinstance(node_content, list):
+                if len(node_schema["prefixItems"]) == 1:
+                    # If there's only one prefix item, this is a single item array, we can just wrap the string
+                    node_content = [node_content]
+                else:
+                    raise TypeError(f"Expected a list or regex for schema node with type array, got {node_content}")
+            if len(node_content) != len(node_schema["prefixItems"]):
+                raise ValueError(
+                    f"Array node has {len(node_content)} items, but schema only has "
+                    f"{len(node_schema['prefixItems'])} prefixItems defined.\n"
+                    f"Content: {node_content}\n"
+                    f"Schema: {node_schema}"
+                )
+            for item, item_schema in zip(node_content, node_schema["prefixItems"]):
+                parsed_schema.append(recursive_parse(item, item_schema))
+            return parsed_schema
+        else:
+            raise ValueError(f"Array node has no items or prefixItems schema defined.\nSchema: {node_schema}")
+    elif node_type in ("string", "integer", "number", "boolean"):
+        if node_type == "integer":
+            if isinstance(node_content, int):
+                return node_content
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Expected a string or int for schema node with type integer, got {type(node_content).__name__}: {node_content}"
+                )
+            try:
+                return int(node_content)
+            except ValueError:
+                raise ValueError(
+                    f"Schema node has type 'integer', but the parsed string content is not a valid integer: {node_content!r}"
+                )
+        elif node_type == "number":
+            if isinstance(node_content, (int, float)):
+                return float(node_content)
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Expected a string or number for schema node with type number, got {type(node_content).__name__}: {node_content}"
+                )
+            try:
+                return float(node_content)
+            except ValueError:
+                raise ValueError(
+                    f"Schema node has type 'number', but the parsed string content is not a valid number: {node_content!r}"
+                )
+        elif node_type == "boolean":
+            if isinstance(node_content, bool):
+                return node_content
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Expected a string or bool for schema node with type boolean, got {type(node_content).__name__}: {node_content}"
+                )
+            if node_content.lower() in ("true", "1"):
+                return True
+            elif node_content.lower() in ("false", "0"):
+                return False
+            else:
+                raise ValueError(f"Invalid boolean value: {node_content}")
+        else:
+            # String type
+            if not isinstance(node_content, str):
+                raise TypeError(
+                    f"Expected a string for schema node with type string, got {type(node_content).__name__}: {node_content}"
+                )
+            return node_content
+    elif node_type is None or node_type == "any":
+        return node_content  # Don't touch it
+    else:
+        raise TypeError(f"Unsupported schema type {node_type} for node: {node_content}")

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/dummy_detectron2_objects.py

@@ -0,0 +1,11 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import requires_backends
+
+
+class LayoutLMv2Model:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["detectron2"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["detectron2"])

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/dummy_music_objects.py

@@ -0,0 +1,16 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class Pop2PianoFeatureExtractor(metaclass=DummyObject):
+    _backends = ["music"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["music"])
+
+
+class Pop2PianoTokenizer(metaclass=DummyObject):
+    _backends = ["music"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["music"])

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/hp_naming.py

@@ -0,0 +1,162 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import re
+
+
+class TrialShortNamer:
+    PREFIX = "hp"
+    DEFAULTS = {}
+    NAMING_INFO = None
+
+    @classmethod
+    def set_defaults(cls, prefix, defaults):
+        cls.PREFIX = prefix
+        cls.DEFAULTS = defaults
+        cls.build_naming_info()
+
+    @staticmethod
+    def shortname_for_word(info, word):
+        if len(word) == 0:
+            return ""
+        short_word = None
+        if any(char.isdigit() for char in word):
+            raise Exception(f"Parameters should not contain numbers: '{word}' contains a number")
+        if word in info["short_word"]:
+            return info["short_word"][word]
+        for prefix_len in range(1, len(word) + 1):
+            prefix = word[:prefix_len]
+            if prefix in info["reverse_short_word"]:
+                continue
+            else:
+                short_word = prefix
+                break
+
+        if short_word is None:
+            # Paranoid fallback
+            def int_to_alphabetic(integer):
+                s = ""
+                while integer != 0:
+                    s = chr(ord("A") + integer % 10) + s
+                    integer //= 10
+                return s
+
+            i = 0
+            while True:
+                sword = word + "#" + int_to_alphabetic(i)
+                if sword in info["reverse_short_word"]:
+                    continue
+                else:
+                    short_word = sword
+                    break
+
+        info["short_word"][word] = short_word
+        info["reverse_short_word"][short_word] = word
+        return short_word
+
+    @staticmethod
+    def shortname_for_key(info, param_name):
+        words = param_name.split("_")
+
+        shortname_parts = [TrialShortNamer.shortname_for_word(info, word) for word in words]
+
+        # We try to create a separatorless short name, but if there is a collision we have to fallback
+        # to a separated short name
+        separators = ["", "_"]
+
+        for separator in separators:
+            shortname = separator.join(shortname_parts)
+            if shortname not in info["reverse_short_param"]:
+                info["short_param"][param_name] = shortname
+                info["reverse_short_param"][shortname] = param_name
+                return shortname
+
+        return param_name
+
+    @staticmethod
+    def add_new_param_name(info, param_name):
+        short_name = TrialShortNamer.shortname_for_key(info, param_name)
+        info["short_param"][param_name] = short_name
+        info["reverse_short_param"][short_name] = param_name
+
+    @classmethod
+    def build_naming_info(cls):
+        if cls.NAMING_INFO is not None:
+            return
+
+        info = {
+            "short_word": {},
+            "reverse_short_word": {},
+            "short_param": {},
+            "reverse_short_param": {},
+        }
+
+        field_keys = list(cls.DEFAULTS.keys())
+
+        for k in field_keys:
+            cls.add_new_param_name(info, k)
+
+        cls.NAMING_INFO = info
+
+    @classmethod
+    def shortname(cls, params):
+        cls.build_naming_info()
+        assert cls.PREFIX is not None
+        name = [copy.copy(cls.PREFIX)]
+
+        for k, v in params.items():
+            if k not in cls.DEFAULTS:
+                raise Exception(f"You should provide a default value for the param name {k} with value {v}")
+            if v == cls.DEFAULTS[k]:
+                # The default value is not added to the name
+                continue
+
+            key = cls.NAMING_INFO["short_param"][k]
+
+            if isinstance(v, bool):
+                v = 1 if v else 0
+
+            sep = "" if isinstance(v, (int, float)) else "-"
+            e = f"{key}{sep}{v}"
+            name.append(e)
+
+        return "_".join(name)
+
+    @classmethod
+    def parse_repr(cls, repr):
+        repr = repr[len(cls.PREFIX) + 1 :]
+        if repr == "":
+            values = []
+        else:
+            values = repr.split("_")
+
+        parameters = {}
+
+        for value in values:
+            if "-" in value:
+                p_k, p_v = value.split("-")
+            else:
+                p_k = re.sub("[0-9.]", "", value)
+                p_v = float(re.sub("[^0-9.]", "", value))
+
+            key = cls.NAMING_INFO["reverse_short_param"][p_k]
+
+            parameters[key] = p_v
+
+        for k in cls.DEFAULTS:
+            if k not in parameters:
+                parameters[k] = cls.DEFAULTS[k]
+
+        return parameters

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/loading_report.py

@@ -0,0 +1,280 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import re
+import shutil
+import sys
+from collections import OrderedDict, defaultdict
+from dataclasses import dataclass
+from typing import Any
+
+
+_DIGIT_RX = re.compile(r"(?<=\.)(\d+)(?=\.|$)")  # numbers between dots or at the end
+
+
+def _pattern_of(key: str) -> str:
+    """Replace every dot-delimited integer with '*' to get the structure."""
+    return _DIGIT_RX.sub("*", key)
+
+
+def _fmt_indices(values: list[int], cutoff=10) -> str:
+    """Format a list of ints as single number, {a, ..., b}, or first...last."""
+    if len(values) == 1:
+        return str(values[0])
+    values = sorted(values)
+    if len(values) > cutoff:
+        return f"{values[0]}...{values[-1]}"
+    return ", ".join(map(str, values))
+
+
+def update_key_name(mapping: dict[str, Any]) -> dict[str, Any]:
+    """
+    Merge keys like 'layers.0.x', 'layers.1.x' into 'layers.{0, 1}.x'
+    BUT only merge together keys that have the exact same value.
+    Returns a new dict {merged_key: value}.
+    """
+    # (pattern, value) -> list[set[int]] (per-star index values)
+    not_mapping = False
+    if not isinstance(mapping, dict):
+        mapping = {k: k for k in mapping}
+        not_mapping = True
+
+    bucket: dict[str, list[set[int] | Any]] = defaultdict(list)
+    for key, val in mapping.items():
+        digs = _DIGIT_RX.findall(key)
+        patt = _pattern_of(key)
+        for i, d in enumerate(digs):
+            if len(bucket[patt]) <= i:
+                bucket[patt].append(set())
+            bucket[patt][i].add(int(d))
+        bucket[patt].append(val)
+
+    out_items = {}
+    for patt, values in bucket.items():
+        sets, val = values[:-1], values[-1]
+        parts = patt.split("*")  # stars are between parts
+        final = parts[0]
+        for i in range(1, len(parts)):
+            if i - 1 < len(sets) and sets[i - 1]:
+                insert = _fmt_indices(sorted(sets[i - 1]))
+                if len(sets[i - 1]) > 1:
+                    final += "{" + insert + "}"
+                else:
+                    final += insert
+            else:
+                final += "*"
+            final += parts[i]
+
+        out_items[final] = val
+    out = OrderedDict(out_items)
+    if not_mapping:
+        return out.keys()
+    return out
+
+
+_ansi_re = re.compile(r"\x1b\[[0-9;]*m")
+
+
+def _strip_ansi(s: str) -> str:
+    return _ansi_re.sub("", str(s))
+
+
+def _pad(text, width):
+    t = str(text)
+    pad = max(0, width - len(_strip_ansi(t)))
+    return t + " " * pad
+
+
+def _make_table(rows, headers):
+    # compute display widths while ignoring ANSI codes
+    cols = list(zip(*([headers] + rows))) if rows else [headers]
+    widths = [max(len(_strip_ansi(x)) for x in col) for col in cols]
+    header_line = " | ".join(_pad(h, w) for h, w in zip(headers, widths))
+    sep_line = "-+-".join("-" * w for w in widths)
+    body = [" | ".join(_pad(c, w) for c, w in zip(r, widths)) for r in rows]
+    return "\n".join([header_line, sep_line] + body)
+
+
+PALETTE = {
+    "reset": "",
+    "red": "",
+    "yellow": "",
+    "orange": "",
+    "purple": "",
+    "bold": "",
+    "italic": "",
+    "dim": "",
+}
+
+
+def _style(s, color):
+    """Return color/style-formatted input `s` if `sys.stdout` is interactive, e.g. connected to a terminal."""
+    if sys.stdout.isatty():
+        return f"{PALETTE[color]}{s}{PALETTE['reset']}"
+    else:
+        return s
+
+
+def _get_terminal_width(default=80):
+    try:
+        return shutil.get_terminal_size().columns
+    except Exception:
+        return default
+
+
+@dataclass
+class LoadStateDictInfo:
+    """
+    Mutable container for state-dict loading results and diagnostics. Each entry in this structure is mutable,
+    and will usually be mutated in-place during the loading pipeline.
+
+    Attributes:
+        missing_keys (`set[str]`):
+            Keys that are missing from the loaded checkpoints but expected in the model's architecture.
+        unexpected_keys (`set[str]`):
+            Keys that are found in the checkpoints, but not expected in the model's architecture.
+        mismatched_keys (`set[tuple[str, tuple[int], tuple[int]]]`):
+            Keys that are found in the checkpoints and are expected in the model's architecture, but with a different shape.
+        error_msgs ( `list[str]`):
+            Some potential error messages.
+        conversion_errors (`dict[str, str]`):
+            Errors happening during the on-the-fly weight conversion process.
+    """
+
+    missing_keys: set[str]
+    unexpected_keys: set[str]
+    mismatched_keys: set[tuple[str, tuple[int], tuple[int]]]
+    error_msgs: list[str]
+    conversion_errors: dict[str, str]
+
+    def missing_and_mismatched(self):
+        """Return all effective missing keys, including `missing` and `mismatched` keys."""
+        return self.missing_keys | {k[0] for k in self.mismatched_keys}
+
+    def to_dict(self):
+        # Does not include the `conversion_errors` to be coherent with legacy reporting in the tests
+        return {
+            "missing_keys": self.missing_keys,
+            "unexpected_keys": self.unexpected_keys,
+            "mismatched_keys": self.mismatched_keys,
+            "error_msgs": self.error_msgs,
+        }
+
+    def create_loading_report(self) -> str | None:
+        """Generate the minimal table of a loading report."""
+        term_w = _get_terminal_width()
+
+        rows = []
+        tips = "\n\nNotes:"
+        if self.unexpected_keys:
+            tips += f"\n- {_style('UNEXPECTED:', 'orange')}\t" + _style(
+                "can be ignored when loading from different task/architecture; not ok if you expect identical arch.",
+                "italic",
+            )
+            for k in update_key_name(self.unexpected_keys):
+                status = _style("UNEXPECTED", "orange")
+                rows.append([k, status, "", ""])
+
+        if self.missing_keys:
+            tips += f"\n- {_style('MISSING:', 'red')}\t" + _style(
+                "those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.",
+                "italic",
+            )
+            for k in update_key_name(self.missing_keys):
+                status = _style("MISSING", "red")
+                rows.append([k, status, ""])
+
+        if self.mismatched_keys:
+            tips += f"\n- {_style('MISMATCH:', 'yellow')}\t" + _style(
+                "ckpt weights were loaded, but they did not match the original empty weight shapes.", "italic"
+            )
+            iterator = {a: (b, c) for a, b, c in self.mismatched_keys}
+            for key, (shape_ckpt, shape_model) in update_key_name(iterator).items():
+                status = _style("MISMATCH", "yellow")
+                data = [
+                    key,
+                    status,
+                    f"Reinit due to size mismatch - ckpt: {str(shape_ckpt)} vs model:{str(shape_model)}",
+                ]
+                rows.append(data)
+
+        if self.conversion_errors:
+            tips += f"\n- {_style('CONVERSION:', 'purple')}\t" + _style(
+                "originate from the conversion scheme", "italic"
+            )
+            for k, v in update_key_name(self.conversion_errors).items():
+                status = _style("CONVERSION", "purple")
+                _details = f"\n\n{v}\n\n"
+                rows.append([k, status, _details])
+
+        # If nothing is wrong, return None
+        if len(rows) == 0:
+            return None
+
+        headers = ["Key", "Status"]
+        if term_w > 200:
+            headers += ["Details"]
+        else:
+            headers += ["", ""]
+        table = _make_table(rows, headers=headers)
+        report = table + tips
+
+        return report
+
+
+def log_state_dict_report(
+    model,
+    pretrained_model_name_or_path: str,
+    ignore_mismatched_sizes: bool,
+    loading_info: LoadStateDictInfo,
+    logger: logging.Logger | None = None,
+):
+    """
+    Log a readable report about state_dict loading issues.
+
+    This version is terminal-size aware: for very small terminals it falls back to a compact
+    Key | Status view so output doesn't wrap badly.
+    """
+    if logger is None:
+        logger = logging.getLogger(__name__)
+
+    # Re-raise errors early if needed
+    if loading_info.error_msgs:
+        error_msg = "\n\t".join(loading_info.error_msgs)
+        if "size mismatch" in error_msg:
+            error_msg += (
+                "\n\tYou may consider adding `ignore_mismatched_sizes=True` to `from_pretrained(...)` if appropriate."
+            )
+        raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
+
+    # Create the report table
+    report = loading_info.create_loading_report()
+    if report is None:
+        return
+
+    prelude = f"{PALETTE['bold']}{model.__class__.__name__} LOAD REPORT{PALETTE['reset']} from: {pretrained_model_name_or_path}\n"
+
+    # Log the report as warning
+    logger.warning(prelude + report)
+
+    # Re-raise in those case, after the report
+    if loading_info.conversion_errors:
+        raise RuntimeError(
+            "We encountered some issues during automatic conversion of the weights. For details look at the `CONVERSION` entries of "
+            "the above report!"
+        )
+    if not ignore_mismatched_sizes and loading_info.mismatched_keys:
+        raise RuntimeError(
+            "You set `ignore_mismatched_sizes` to `False`, thus raising an error. For details look at the above report!"
+        )

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/logging.py

@@ -0,0 +1,441 @@
+# Copyright 2020 Optuna, Hugging Face
+#
+# 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.
+"""Logging utilities."""
+
+import functools
+import logging
+import os
+import sys
+import threading
+from collections.abc import Callable
+from logging import (
+    CRITICAL,  # NOQA
+    DEBUG,
+    ERROR,
+    FATAL,  # NOQA
+    INFO,
+    NOTSET,  # NOQA
+    WARN,  # NOQA
+    WARNING,
+)
+from logging import captureWarnings as _captureWarnings
+from typing import Any
+
+import huggingface_hub.utils as hf_hub_utils
+from tqdm import auto as tqdm_lib
+
+from .._typing import TransformersLogger
+
+
+_lock = threading.Lock()
+_default_handler: logging.Handler | None = None
+
+log_levels = {
+    "detail": logging.DEBUG,  # will also print filename and line number
+    "debug": logging.DEBUG,
+    "info": logging.INFO,
+    "warning": logging.WARNING,
+    "error": logging.ERROR,
+    "critical": logging.CRITICAL,
+}
+
+_default_log_level = logging.WARNING
+
+_tqdm_active = not hf_hub_utils.are_progress_bars_disabled()
+_tqdm_hook: Callable[[Callable[..., Any], tuple[Any, ...], dict[str, Any]], Any] | None = None
+
+
+def _get_default_logging_level():
+    """
+    If TRANSFORMERS_VERBOSITY env var is set to one of the valid choices return that as the new default level. If it is
+    not - fall back to `_default_log_level`
+    """
+    env_level_str = os.getenv("TRANSFORMERS_VERBOSITY", None)
+    if env_level_str:
+        if env_level_str in log_levels:
+            return log_levels[env_level_str]
+        else:
+            logging.getLogger().warning(
+                f"Unknown option TRANSFORMERS_VERBOSITY={env_level_str}, "
+                f"has to be one of: {', '.join(log_levels.keys())}"
+            )
+    return _default_log_level
+
+
+def _get_library_name() -> str:
+    return __name__.split(".")[0]
+
+
+def _get_library_root_logger() -> logging.Logger:
+    return logging.getLogger(_get_library_name())
+
+
+def _configure_library_root_logger() -> None:
+    global _default_handler
+
+    with _lock:
+        if _default_handler:
+            # This library has already configured the library root logger.
+            return
+        _default_handler = logging.StreamHandler()  # Set sys.stderr as stream.
+        # set defaults based on https://github.com/pyinstaller/pyinstaller/issues/7334#issuecomment-1357447176
+        if sys.stderr is None:
+            sys.stderr = open(os.devnull, "w")
+
+        _default_handler.flush = sys.stderr.flush
+
+        # Apply our default configuration to the library root logger.
+        library_root_logger = _get_library_root_logger()
+        library_root_logger.addHandler(_default_handler)
+        library_root_logger.setLevel(_get_default_logging_level())
+        # Always show lib when logging in non-verbose mode. Note, other libs
+        # use `transformers.logger` directly, so we check `lib_name` to be safe
+        lib_name = _get_library_name()
+        logging_format = f"[{lib_name}] %(message)s"
+
+        # if logging level is debug, we add pathname and lineno to formatter for easy debugging
+        if os.getenv("TRANSFORMERS_VERBOSITY", None) == "detail":
+            logging_format = "%(levelname)s [%(name)s:%(lineno)s] %(asctime)s %(message)s"
+
+        formatter = logging.Formatter(logging_format)
+        _default_handler.setFormatter(formatter)
+
+        ci = os.getenv("CI")
+        is_ci = ci is not None and ci.upper() in {"1", "ON", "YES", "TRUE"}
+        library_root_logger.propagate = is_ci
+
+
+def _reset_library_root_logger() -> None:
+    global _default_handler
+
+    with _lock:
+        if not _default_handler:
+            return
+
+        library_root_logger = _get_library_root_logger()
+        library_root_logger.removeHandler(_default_handler)
+        library_root_logger.setLevel(logging.NOTSET)
+        _default_handler = None
+
+
+def get_log_levels_dict():
+    return log_levels
+
+
+def captureWarnings(capture):
+    """
+    Calls the `captureWarnings` method from the logging library to enable management of the warnings emitted by the
+    `warnings` library.
+
+    Read more about this method here:
+    https://docs.python.org/3/library/logging.html#integration-with-the-warnings-module
+
+    All warnings will be logged through the `py.warnings` logger.
+
+    Careful: this method also adds a handler to this logger if it does not already have one, and updates the logging
+    level of that logger to the library's root logger.
+    """
+    logger = get_logger("py.warnings")
+
+    if not logger.handlers:
+        logger.addHandler(_default_handler)
+
+    logger.setLevel(_get_library_root_logger().level)
+
+    _captureWarnings(capture)
+
+
+def get_logger(name: str | None = None) -> TransformersLogger:
+    """
+    Return a logger with the specified name.
+
+    This function is not supposed to be directly accessed unless you are writing a custom transformers module.
+    """
+
+    if name is None:
+        name = _get_library_name()
+
+    _configure_library_root_logger()
+    return logging.getLogger(name)
+
+
+def get_verbosity() -> int:
+    """
+    Return the current level for the 🤗 Transformers's root logger as an int.
+
+    Returns:
+        `int`: The logging level.
+
+    <Tip>
+
+    🤗 Transformers has following logging levels:
+
+    - 50: `transformers.logging.CRITICAL` or `transformers.logging.FATAL`
+    - 40: `transformers.logging.ERROR`
+    - 30: `transformers.logging.WARNING` or `transformers.logging.WARN`
+    - 20: `transformers.logging.INFO`
+    - 10: `transformers.logging.DEBUG`
+
+    </Tip>"""
+
+    _configure_library_root_logger()
+    return _get_library_root_logger().getEffectiveLevel()
+
+
+def set_verbosity(verbosity: int) -> None:
+    """
+    Set the verbosity level for the 🤗 Transformers's root logger.
+
+    Args:
+        verbosity (`int`):
+            Logging level, e.g., one of:
+
+            - `transformers.logging.CRITICAL` or `transformers.logging.FATAL`
+            - `transformers.logging.ERROR`
+            - `transformers.logging.WARNING` or `transformers.logging.WARN`
+            - `transformers.logging.INFO`
+            - `transformers.logging.DEBUG`
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().setLevel(verbosity)
+
+
+def set_verbosity_info():
+    """Set the verbosity to the `INFO` level."""
+    return set_verbosity(INFO)
+
+
+def set_verbosity_warning():
+    """Set the verbosity to the `WARNING` level."""
+    return set_verbosity(WARNING)
+
+
+def set_verbosity_debug():
+    """Set the verbosity to the `DEBUG` level."""
+    return set_verbosity(DEBUG)
+
+
+def set_verbosity_error():
+    """Set the verbosity to the `ERROR` level."""
+    return set_verbosity(ERROR)
+
+
+def disable_default_handler() -> None:
+    """Disable the default handler of the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert _default_handler is not None
+    _get_library_root_logger().removeHandler(_default_handler)
+
+
+def enable_default_handler() -> None:
+    """Enable the default handler of the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert _default_handler is not None
+    _get_library_root_logger().addHandler(_default_handler)
+
+
+def add_handler(handler: logging.Handler) -> None:
+    """adds a handler to the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert handler is not None
+    _get_library_root_logger().addHandler(handler)
+
+
+def remove_handler(handler: logging.Handler) -> None:
+    """removes given handler from the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert handler is not None and handler not in _get_library_root_logger().handlers
+    _get_library_root_logger().removeHandler(handler)
+
+
+def disable_propagation() -> None:
+    """
+    Disable propagation of the library log outputs. Note that log propagation is disabled by default.
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().propagate = False
+
+
+def enable_propagation() -> None:
+    """
+    Enable propagation of the library log outputs. Please disable the HuggingFace Transformers's default handler to
+    prevent double logging if the root logger has been configured.
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().propagate = True
+
+
+def enable_explicit_format() -> None:
+    """
+    Enable explicit formatting for every HuggingFace Transformers's logger. The explicit formatter is as follows:
+    ```
+        [LEVELNAME|FILENAME|LINE NUMBER] TIME >> MESSAGE
+    ```
+    All handlers currently bound to the root logger are affected by this method.
+    """
+    handlers = _get_library_root_logger().handlers
+
+    for handler in handlers:
+        formatter = logging.Formatter("[%(levelname)s|%(filename)s:%(lineno)s] %(asctime)s >> %(message)s")
+        handler.setFormatter(formatter)
+
+
+def reset_format() -> None:
+    """
+    Resets the formatting for HuggingFace Transformers's loggers.
+
+    All handlers currently bound to the root logger are affected by this method.
+    """
+    handlers = _get_library_root_logger().handlers
+
+    for handler in handlers:
+        handler.setFormatter(None)
+
+
+def warning_advice(self, *args, **kwargs):
+    """
+    This method is identical to `logger.warning()`, but if env var TRANSFORMERS_NO_ADVISORY_WARNINGS=1 is set, this
+    warning will not be printed
+    """
+    no_advisory_warnings = os.getenv("TRANSFORMERS_NO_ADVISORY_WARNINGS")
+    if no_advisory_warnings:
+        return
+    self.warning(*args, **kwargs)
+
+
+logging.Logger.warning_advice = warning_advice  # type: ignore[unresolved-attribute]
+
+
+@functools.lru_cache(None)
+def warning_once(self, *args, **kwargs):
+    """
+    This method is identical to `logger.warning()`, but will emit the warning with the same message only once
+
+    Note: The cache is for the function arguments, so 2 different callers using the same arguments will hit the cache.
+    The assumption here is that all warning messages are unique across the code. If they aren't then need to switch to
+    another type of cache that includes the caller frame information in the hashing function.
+    """
+    self.warning(*args, **kwargs)
+
+
+logging.Logger.warning_once = warning_once  # type: ignore[unresolved-attribute]
+
+
+@functools.lru_cache(None)
+def info_once(self, *args, **kwargs):
+    """
+    This method is identical to `logger.info()`, but will emit the info with the same message only once
+
+    Note: The cache is for the function arguments, so 2 different callers using the same arguments will hit the cache.
+    The assumption here is that all warning messages are unique across the code. If they aren't then need to switch to
+    another type of cache that includes the caller frame information in the hashing function.
+    """
+    self.info(*args, **kwargs)
+
+
+logging.Logger.info_once = info_once  # type: ignore[unresolved-attribute]
+
+
+class EmptyTqdm:
+    """Dummy tqdm which doesn't do anything."""
+
+    def __init__(self, *args, **kwargs):  # pylint: disable=unused-argument
+        self._iterator = args[0] if args else None
+
+    def __iter__(self):
+        return iter(self._iterator)
+
+    def __getattr__(self, _):
+        """Return empty function."""
+
+        def empty_fn(*args, **kwargs):  # pylint: disable=unused-argument
+            return
+
+        return empty_fn
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, type_, value, traceback):
+        return
+
+
+class _tqdm_cls:
+    def __call__(self, *args, **kwargs):
+        factory = tqdm_lib.tqdm if _tqdm_active else EmptyTqdm
+        if _tqdm_hook is not None:
+            return _tqdm_hook(factory, args, kwargs)
+        return factory(*args, **kwargs)
+
+    def set_lock(self, *args, **kwargs):
+        self._lock = None
+        if _tqdm_active:
+            return tqdm_lib.tqdm.set_lock(*args, **kwargs)
+
+    def get_lock(self):
+        if _tqdm_active:
+            return tqdm_lib.tqdm.get_lock()
+
+
+tqdm = _tqdm_cls()
+
+
+def is_progress_bar_enabled() -> bool:
+    """Return a boolean indicating whether tqdm progress bars are enabled."""
+    return bool(_tqdm_active)
+
+
+def enable_progress_bar():
+    """Enable tqdm progress bar."""
+    global _tqdm_active
+    _tqdm_active = True
+    hf_hub_utils.enable_progress_bars()
+
+
+def disable_progress_bar():
+    """Disable tqdm progress bar."""
+    global _tqdm_active
+    _tqdm_active = False
+    hf_hub_utils.disable_progress_bars()
+
+
+def set_tqdm_hook(hook: Callable[[Callable[..., Any], tuple[Any, ...], dict[str, Any]], Any] | None):
+    """
+    Set a hook that customizes tqdm creation.
+
+    The hook is called with the tqdm factory to use (either `tqdm.auto.tqdm` or an empty shim), along with the
+    positional and keyword arguments that would have been passed to tqdm. The hook should return an object compatible
+    with tqdm (i.e. implementing the methods your code relies on, such as `update`, `close`, context manager methods,
+    etc.).
+
+    Passing `None` clears the hook.
+
+    Returns:
+        The previous hook, which can be restored later.
+    """
+    global _tqdm_hook
+    previous_hook = _tqdm_hook
+    _tqdm_hook = hook
+    return previous_hook

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/network_logging.py

@@ -0,0 +1,485 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import inspect
+import json
+import os
+import threading
+import time
+from collections import defaultdict
+from functools import wraps
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from .generic import strtobool
+
+
+class _NetworkRequestTrace:
+    def __init__(self, request: httpx.Request):
+        self.request = request
+        self.started_at = time.perf_counter()
+        self.phase_started_at = {}
+        self.phases_ms = defaultdict(float)
+
+    def trace(self, name: str, info: dict[str, Any]) -> None:
+        parts = name.rsplit(".", 2)
+        if len(parts) != 3:
+            return
+
+        _, phase, state = parts
+        now = time.perf_counter()
+        if state == "started":
+            self.phase_started_at[phase] = now
+        elif state in {"complete", "failed"}:
+            phase_started_at = self.phase_started_at.pop(phase, None)
+            if phase_started_at is not None:
+                self.phases_ms[phase] += (now - phase_started_at) * 1000
+
+    def build_record(
+        self,
+        *,
+        response: httpx.Response | None = None,
+        error: BaseException | None = None,
+        stream: bool = False,
+    ) -> dict[str, Any]:
+        total_ms = (time.perf_counter() - self.started_at) * 1000
+        url = self.request.url
+        host = url.host or ""
+        port = url.port
+        default_port = {"http": 80, "https": 443}.get(url.scheme)
+        host_display = host if port in (None, default_port) else f"{host}:{port}"
+
+        http_version = None
+        status_code = None
+        bytes_downloaded = None
+        response_complete = False
+        if response is not None:
+            status_code = response.status_code
+            response_complete = response.is_closed
+            raw_http_version = response.extensions.get("http_version")
+            if isinstance(raw_http_version, bytes):
+                http_version = raw_http_version.decode("ascii", errors="replace")
+            elif raw_http_version is not None:
+                http_version = str(raw_http_version)
+
+            if response_complete:
+                try:
+                    bytes_downloaded = len(response.content)
+                except httpx.ResponseNotRead:
+                    pass
+
+        return {
+            "method": self.request.method,
+            "scheme": url.scheme,
+            "host": host,
+            "host_display": host_display,
+            "port": port,
+            "path": url.path,
+            "has_query": bool(url.query),
+            "url": f"{url.scheme}://{host_display}{url.path}{'?...' if url.query else ''}",
+            "request_id": self.request.headers.get("x-amzn-trace-id") or self.request.headers.get("x-request-id"),
+            "status_code": status_code,
+            "http_version": http_version,
+            "bytes_downloaded": bytes_downloaded,
+            "total_ms": total_ms,
+            "stream": stream,
+            "response_complete": response_complete,
+            "phases_ms": dict(sorted(self.phases_ms.items())),
+            "error": None if error is None else f"{type(error).__name__}: {error}",
+        }
+
+
+class _NetworkDebugProfiler:
+    def __init__(self):
+        self._records = []
+        self._lock = threading.Lock()
+        self._enabled = False
+        self._output_path = None
+        self._original_client_send = None
+        self._original_async_client_send = None
+        self._shared_dir = None
+
+    @property
+    def enabled(self) -> bool:
+        return self._enabled
+
+    def clear(self) -> None:
+        with self._lock:
+            self._records = []
+
+    def enable(self, output_path: str | os.PathLike | None = None) -> None:
+        if self._enabled:
+            self._output_path = None if output_path is None else os.fspath(output_path)
+            self.clear()
+            return
+
+        self._output_path = None if output_path is None else os.fspath(output_path)
+        self.clear()
+
+        profiler = self
+        self._original_client_send = httpx.Client.send
+        self._original_async_client_send = httpx.AsyncClient.send
+
+        @wraps(self._original_client_send)
+        def patched_client_send(client, request, *args, **kwargs):
+            return profiler._send_with_trace(profiler._original_client_send, client, request, *args, **kwargs)
+
+        @wraps(self._original_async_client_send)
+        async def patched_async_client_send(client, request, *args, **kwargs):
+            return await profiler._async_send_with_trace(
+                profiler._original_async_client_send, client, request, *args, **kwargs
+            )
+
+        httpx.Client.send = patched_client_send
+        httpx.AsyncClient.send = patched_async_client_send
+        self._enabled = True
+
+    def setup_shared_dir(self) -> str | None:
+        """Create a shared temp directory for xdist workers to dump records into."""
+        if self._shared_dir is None:
+            import tempfile
+
+            self._shared_dir = tempfile.mkdtemp(prefix="network_debug_")
+        return self._shared_dir
+
+    def set_shared_dir(self, shared_dir: str) -> None:
+        """Set the shared directory (called in xdist workers)."""
+        self._shared_dir = shared_dir
+
+    def dump_worker_records(self, worker_id: str | None = None) -> None:
+        """Write this process's records to a file in the shared directory (called in workers)."""
+        if not self._shared_dir or not self._records:
+            return
+        worker_id = worker_id or f"pid{os.getpid()}"
+        dump_path = os.path.join(self._shared_dir, f"records_{worker_id}.json")
+        with self._lock:
+            records = [{**record, "phases_ms": dict(record["phases_ms"])} for record in self._records]
+        Path(dump_path).write_text(json.dumps(records), encoding="utf-8")
+
+    def load_worker_records(self) -> None:
+        """Load all worker record files from the shared directory (called in controller)."""
+        if not self._shared_dir or not os.path.isdir(self._shared_dir):
+            return
+        import glob as glob_module
+
+        for record_file in glob_module.glob(os.path.join(self._shared_dir, "records_*.json")):
+            try:
+                records = json.loads(Path(record_file).read_text(encoding="utf-8"))
+                with self._lock:
+                    for record in records:
+                        record["phases_ms"] = defaultdict(float, record.get("phases_ms", {}))
+                        self._records.append(record)
+            except (OSError, json.JSONDecodeError):
+                pass
+
+    def cleanup_shared_dir(self) -> None:
+        """Remove the shared temp directory."""
+        if self._shared_dir and os.path.isdir(self._shared_dir):
+            import shutil
+
+            shutil.rmtree(self._shared_dir, ignore_errors=True)
+            self._shared_dir = None
+
+    def disable(self) -> None:
+        if not self._enabled:
+            return
+
+        httpx.Client.send = self._original_client_send
+        httpx.AsyncClient.send = self._original_async_client_send
+        self._enabled = False
+        self._original_client_send = None
+        self._original_async_client_send = None
+        self._output_path = None
+        self.clear()
+
+    def _append_record(self, record: dict[str, Any]) -> None:
+        with self._lock:
+            self._records.append(record)
+
+    def _wrap_trace_callback(self, request: httpx.Request, trace: _NetworkRequestTrace):
+        existing_trace = request.extensions.get("trace")
+
+        def wrapped_trace(name: str, info: dict[str, Any]) -> Any:
+            trace.trace(name, info)
+            if existing_trace is not None:
+                return existing_trace(name, info)
+            return None
+
+        return wrapped_trace
+
+    async def _awrap_trace_callback(self, request: httpx.Request, trace: _NetworkRequestTrace):
+        existing_trace = request.extensions.get("trace")
+
+        async def wrapped_trace(name: str, info: dict[str, Any]) -> Any:
+            trace.trace(name, info)
+            if existing_trace is not None:
+                result = existing_trace(name, info)
+                if inspect.isawaitable(result):
+                    return await result
+                return result
+            return None
+
+        return wrapped_trace
+
+    def _send_with_trace(self, original_send, client, request: httpx.Request, *args, **kwargs):
+        trace = _NetworkRequestTrace(request)
+        request.extensions = dict(request.extensions)
+        request.extensions["trace"] = self._wrap_trace_callback(request, trace)
+
+        try:
+            response = original_send(client, request, *args, **kwargs)
+        except Exception as error:
+            self._append_record(trace.build_record(error=error, stream=kwargs.get("stream", False)))
+            raise
+
+        self._append_record(trace.build_record(response=response, stream=kwargs.get("stream", False)))
+        return response
+
+    async def _async_send_with_trace(self, original_send, client, request: httpx.Request, *args, **kwargs):
+        trace = _NetworkRequestTrace(request)
+        request.extensions = dict(request.extensions)
+        request.extensions["trace"] = await self._awrap_trace_callback(request, trace)
+
+        try:
+            response = await original_send(client, request, *args, **kwargs)
+        except Exception as error:
+            self._append_record(trace.build_record(error=error, stream=kwargs.get("stream", False)))
+            raise
+
+        self._append_record(trace.build_record(response=response, stream=kwargs.get("stream", False)))
+        return response
+
+    def build_report(self) -> dict[str, Any]:
+        with self._lock:
+            records = [
+                {
+                    **record,
+                    "phases_ms": dict(record["phases_ms"]),
+                }
+                for record in self._records
+            ]
+
+        phase_totals_ms = defaultdict(float)
+        route_totals = {}
+        for record in records:
+            for phase, duration_ms in record["phases_ms"].items():
+                phase_totals_ms[phase] += duration_ms
+
+            route_key = (record["method"], record["host_display"], record["path"])
+            route_total = route_totals.setdefault(
+                route_key,
+                {
+                    "method": record["method"],
+                    "host_display": record["host_display"],
+                    "path": record["path"],
+                    "count": 0,
+                    "failures": 0,
+                    "total_ms": 0.0,
+                    "phase_totals_ms": defaultdict(float),
+                },
+            )
+            route_total["count"] += 1
+            route_total["total_ms"] += record["total_ms"]
+            route_total["failures"] += int(record["error"] is not None)
+            for phase, duration_ms in record["phases_ms"].items():
+                route_total["phase_totals_ms"][phase] += duration_ms
+
+        routes = []
+        for route_total in route_totals.values():
+            route_total["avg_ms"] = route_total["total_ms"] / route_total["count"]
+            route_total["phase_totals_ms"] = dict(sorted(route_total["phase_totals_ms"].items()))
+            routes.append(route_total)
+
+        routes.sort(key=lambda route: route["total_ms"], reverse=True)
+        total_time_ms = sum(record["total_ms"] for record in records)
+        return {
+            "enabled": self._enabled,
+            "output_path": self._output_path,
+            "total_requests": len(records),
+            "failed_requests": sum(int(record["error"] is not None) for record in records),
+            "total_time_ms": total_time_ms,
+            "phase_totals_ms": dict(sorted(phase_totals_ms.items())),
+            "requests": records,
+            "routes": routes,
+        }
+
+    def maybe_write_report(self) -> str | None:
+        if self._output_path is None:
+            return None
+
+        report_path = Path(self._output_path)
+        report_path.parent.mkdir(parents=True, exist_ok=True)
+        report_path.write_text(json.dumps(self.build_report(), indent=2, sort_keys=True), encoding="utf-8")
+        return str(report_path)
+
+
+_NETWORK_DEBUG_PROFILER = _NetworkDebugProfiler()
+
+
+_DEFAULT_REPORT_PATH = "network_debug_report.json"
+
+
+def _parse_network_debug_env() -> tuple[bool, str]:
+    enabled_raw = os.environ.get("NETWORK_DEBUG_REPORT", "").strip()
+    try:
+        enabled = bool(strtobool(enabled_raw)) if enabled_raw else False
+    except ValueError:
+        enabled = False
+
+    output_path = os.environ.get("NETWORK_DEBUG_REPORT_PATH", "").strip() or _DEFAULT_REPORT_PATH
+    return enabled, output_path
+
+
+def _enable_network_debug_report(output_path: str | os.PathLike | None = None) -> None:
+    _NETWORK_DEBUG_PROFILER.enable(output_path=output_path)
+
+
+def _disable_network_debug_report() -> None:
+    _NETWORK_DEBUG_PROFILER.disable()
+
+
+def _clear_network_debug_report() -> None:
+    _NETWORK_DEBUG_PROFILER.clear()
+
+
+def _get_network_debug_report() -> dict[str, Any]:
+    return _NETWORK_DEBUG_PROFILER.build_report()
+
+
+def _enable_network_debug_report_from_env() -> bool:
+    enabled, output_path = _parse_network_debug_env()
+    if not enabled:
+        return False
+
+    _enable_network_debug_report(output_path=output_path)
+    return True
+
+
+def _format_network_debug_report(max_requests: int = 20, max_routes: int = 10) -> str:
+    report = _get_network_debug_report()
+    if report["total_requests"] == 0:
+        return "Network debug report: no httpx requests captured."
+
+    lines = [
+        "Network debug report",
+        f"Requests captured: {report['total_requests']}",
+        f"Failed requests: {report['failed_requests']}",
+        f"Cumulative request time: {report['total_time_ms']:.1f} ms",
+    ]
+
+    if report["phase_totals_ms"]:
+        phase_summary = ", ".join(
+            f"{phase}={duration_ms:.1f} ms"
+            for phase, duration_ms in sorted(report["phase_totals_ms"].items(), key=lambda item: item[1], reverse=True)
+        )
+        lines.append(f"Phase totals: {phase_summary}")
+
+    lines.append("")
+    lines.append("Slowest requests:")
+    for idx, record in enumerate(
+        sorted(report["requests"], key=lambda request: request["total_ms"], reverse=True)[:max_requests],
+        start=1,
+    ):
+        status = record["error"] or f"status={record['status_code']}"
+        phase_bits = []
+        for phase in ("connect_tcp", "start_tls", "receive_response_headers", "receive_response_body"):
+            duration_ms = record["phases_ms"].get(phase)
+            if duration_ms is not None:
+                phase_bits.append(f"{phase}={duration_ms:.1f} ms")
+        phase_suffix = f" ({', '.join(phase_bits)})" if phase_bits else ""
+        incomplete_suffix = " incomplete" if record["stream"] and not record["response_complete"] else ""
+        lines.append(
+            f"{idx:>2}. {record['method']} {record['url']} {record['total_ms']:.1f} ms {status}{incomplete_suffix}{phase_suffix}"
+        )
+
+    lines.append("")
+    lines.append("Slowest routes:")
+    for idx, route in enumerate(report["routes"][:max_routes], start=1):
+        lines.append(
+            f"{idx:>2}. {route['method']} {route['host_display']}{route['path']} count={route['count']} "
+            f"total={route['total_ms']:.1f} ms avg={route['avg_ms']:.1f} ms failures={route['failures']}"
+        )
+
+    return "\n".join(lines)
+
+
+class NetworkDebugPlugin:
+    """Pytest plugin that handles all network debug orchestration including xdist coordination."""
+
+    def pytest_configure(self, config):
+        _enable_network_debug_report_from_env()
+        if not _NETWORK_DEBUG_PROFILER.enabled:
+            return
+
+        # xdist controller: create shared dir for workers to dump network records
+        if not hasattr(config, "workerinput"):
+            shared_dir = _NETWORK_DEBUG_PROFILER.setup_shared_dir()
+            if shared_dir:
+                config._network_debug_shared_dir = shared_dir
+        else:
+            # xdist worker: receive shared dir from controller
+            shared_dir = config.workerinput.get("network_debug_shared_dir")
+            if shared_dir:
+                _NETWORK_DEBUG_PROFILER.set_shared_dir(shared_dir)
+
+    def pytest_configure_node(self, node):
+        """xdist hook: called on the controller to configure each worker node."""
+        shared_dir = getattr(node.config, "_network_debug_shared_dir", None)
+        if shared_dir:
+            node.workerinput["network_debug_shared_dir"] = shared_dir
+
+    def pytest_sessionfinish(self, session, exitstatus):
+        # xdist worker: dump network debug records for the controller to aggregate
+        if hasattr(session.config, "workerinput"):
+            worker_id = session.config.workerinput.get("workerid", f"pid{os.getpid()}")
+            _NETWORK_DEBUG_PROFILER.dump_worker_records(worker_id=worker_id)
+
+    def pytest_terminal_summary(self, terminalreporter):
+        if not _NETWORK_DEBUG_PROFILER.enabled:
+            return
+
+        # Skip report generation in xdist worker processes; only the controller should aggregate and report.
+        if hasattr(terminalreporter.config, "workerinput"):
+            return
+
+        # Aggregate worker records if running under xdist.
+        _NETWORK_DEBUG_PROFILER.load_worker_records()
+
+        report_path = None
+        try:
+            report_path = _NETWORK_DEBUG_PROFILER.maybe_write_report()
+        except OSError as error:
+            report_path = f"Failed to write JSON report: {error}"
+
+        terminalreporter.section("Network debug", sep="=")
+        for line in _format_network_debug_report().splitlines():
+            terminalreporter.write_line(line)
+        if report_path is not None:
+            terminalreporter.write_line(f"JSON report: {report_path}")
+
+        _NETWORK_DEBUG_PROFILER.cleanup_shared_dir()
+
+
+def register_network_debug_plugin(config) -> None:
+    """Register the network debug pytest plugin. Single entry point for conftest.py."""
+    config.pluginmanager.register(NetworkDebugPlugin(), "network_debug")
+
+
+__all__ = [
+    "register_network_debug_plugin",
+]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/peft_utils.py

@@ -0,0 +1,117 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import importlib.metadata
+import os
+
+from packaging import version
+
+from .hub import cached_file
+from .import_utils import is_peft_available
+
+
+ADAPTER_CONFIG_NAME = "adapter_config.json"
+ADAPTER_WEIGHTS_NAME = "adapter_model.bin"
+ADAPTER_SAFE_WEIGHTS_NAME = "adapter_model.safetensors"
+
+
+def find_adapter_config_file(
+    model_id: str,
+    cache_dir: str | os.PathLike | None = None,
+    force_download: bool = False,
+    proxies: dict[str, str] | None = None,
+    token: bool | str | None = None,
+    revision: str | None = None,
+    local_files_only: bool = False,
+    subfolder: str = "",
+    _commit_hash: str | None = None,
+) -> str | None:
+    r"""
+    Simply checks if the model stored on the Hub or locally is an adapter model or not, return the path of the adapter
+    config file if it is, None otherwise.
+
+    Args:
+        model_id (`str`):
+            The identifier of the model to look for, can be either a local path or an id to the repository on the Hub.
+        cache_dir (`str` or `os.PathLike`, *optional*):
+            Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+            cache should not be used.
+        force_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to force to (re-)download the configuration files and override the cached versions if they
+            exist.
+        proxies (`dict[str, str]`, *optional*):
+            A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+            'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+        token (`str` or *bool*, *optional*):
+            The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+            when running `hf auth login` (stored in `~/.huggingface`).
+        revision (`str`, *optional*, defaults to `"main"`):
+            The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+            git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+            identifier allowed by git.
+
+            <Tip>
+
+            To test a pull request you made on the Hub, you can pass `revision="refs/pr/<pr_number>".
+
+            </Tip>
+
+        local_files_only (`bool`, *optional*, defaults to `False`):
+            If `True`, will only try to load the tokenizer configuration from local files.
+        subfolder (`str`, *optional*, defaults to `""`):
+            In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can
+            specify the folder name here.
+    """
+    adapter_cached_filename = None
+    if model_id is None:
+        return None
+    elif os.path.isdir(model_id):
+        list_remote_files = os.listdir(model_id)
+        if ADAPTER_CONFIG_NAME in list_remote_files:
+            adapter_cached_filename = os.path.join(model_id, ADAPTER_CONFIG_NAME)
+    else:
+        adapter_cached_filename = cached_file(
+            model_id,
+            ADAPTER_CONFIG_NAME,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
+            subfolder=subfolder,
+            _commit_hash=_commit_hash,
+            _raise_exceptions_for_gated_repo=False,
+            _raise_exceptions_for_missing_entries=False,
+            _raise_exceptions_for_connection_errors=False,
+        )
+
+    return adapter_cached_filename
+
+
+def check_peft_version(min_version: str) -> None:
+    r"""
+    Checks if the version of PEFT is compatible.
+
+    Args:
+        version (`str`):
+            The version of PEFT to check against.
+    """
+    if not is_peft_available():
+        raise ValueError("PEFT is not installed. Please install it with `pip install peft`")
+
+    is_peft_version_compatible = version.parse(importlib.metadata.version("peft")) >= version.parse(min_version)
+
+    if not is_peft_version_compatible:
+        raise ValueError(f"The version of PEFT you are using is not compatible, please use a version >= {min_version}")

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/utils/versions.py

@@ -0,0 +1,116 @@
+# Copyright 2020 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.
+"""
+Utilities for working with package versions
+"""
+
+import importlib.metadata
+import operator
+import re
+import sys
+
+from packaging import version
+
+
+ops = {
+    "<": operator.lt,
+    "<=": operator.le,
+    "==": operator.eq,
+    "!=": operator.ne,
+    ">=": operator.ge,
+    ">": operator.gt,
+}
+
+
+def _compare_versions(op, got_ver, want_ver, requirement, pkg, hint):
+    if got_ver is None or want_ver is None:
+        raise ValueError(
+            f"Unable to compare versions for {requirement}: need={want_ver} found={got_ver}. This is unusual. Consider"
+            f" reinstalling {pkg}."
+        )
+    if not ops[op](version.parse(got_ver), version.parse(want_ver)):
+        raise ImportError(
+            f"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}"
+        )
+
+
+def require_version(requirement: str, hint: str | None = None) -> None:
+    """
+    Perform a runtime check of the dependency versions, using the exact same syntax used by pip.
+
+    The installed module version comes from the *site-packages* dir via *importlib.metadata*.
+
+    Args:
+        requirement (`str`): pip style definition, e.g.,  "tokenizers==0.9.4", "tqdm>=4.27", "numpy"
+        hint (`str`, *optional*): what suggestion to print in case of requirements not being met
+
+    Example:
+
+    ```python
+    require_version("pandas>1.1.2")
+    require_version("numpy>1.18.5", "this is important to have for whatever reason")
+    ```"""
+
+    hint = f"\n{hint}" if hint is not None else ""
+
+    # non-versioned check
+    if re.match(r"^[\w_\-\d]+$", requirement):
+        pkg, op, want_ver = requirement, None, None
+    else:
+        match = re.findall(r"^([^!=<>\s]+)([\s!=<>]{1,2}.+)", requirement)
+        if not match:
+            raise ValueError(
+                "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but"
+                f" got {requirement}"
+            )
+        pkg, want_full = match[0]
+        want_range = want_full.split(",")  # there could be multiple requirements
+        wanted = {}
+        for w in want_range:
+            match = re.findall(r"^([\s!=<>]{1,2})(.+)", w)
+            if not match:
+                raise ValueError(
+                    "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23,"
+                    f" but got {requirement}"
+                )
+            op, want_ver = match[0]
+            wanted[op] = want_ver
+            if op not in ops:
+                raise ValueError(f"{requirement}: need one of {list(ops.keys())}, but got {op}")
+
+    # special case
+    if pkg == "python":
+        got_ver = ".".join([str(x) for x in sys.version_info[:3]])
+        for op, want_ver in wanted.items():
+            _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+        return
+
+    # check if any version is installed
+    try:
+        got_ver = importlib.metadata.version(pkg)
+    except importlib.metadata.PackageNotFoundError:
+        raise importlib.metadata.PackageNotFoundError(
+            f"The '{requirement}' distribution was not found and is required by this application. {hint}"
+        )
+
+    # check that the right version is installed if version number or a range was provided
+    if want_ver is not None:
+        for op, want_ver in wanted.items():
+            _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+
+
+def require_version_core(requirement):
+    """require_version wrapper which emits a core-specific hint on failure"""
+    hint = "Try: `pip install transformers -U` or `pip install -e '.[dev]'` if you're working with git main"
+    return require_version(requirement, hint)