[Model] Add PP-DocLayoutV3 Model Support by zhang-prog · Pull Request #43098 · huggingface/transformers

zhang-prog · 2026-01-04T12:10:22Z

No description provided.

ArthurZucker

hey! can you provide a bit of contexte for us? 🤗 a link to the model being released!
Also the code looks very far aways from https://huggingface.co/docs/transformers/v4.48.0/modular_transformers which I inviite you to read to get a better idea of how to adapt the code!

zhang-prog · 2026-01-07T08:41:27Z

@ArthurZucker

The model is not yet publicly released, but we aim to merge it into the transformers library as soon as possible.
I’ve adopted a modular approach and further optimized the code.

PP-DocLayoutV3 model_doc and huggingface repo are comming soon. Could I please request a review?

molbap

Hello, looking forward to see this out! Wrote a first review, don't forget to fill the toctree and add the model documentation as well! Ping me when done and I'll re-review 🤗

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

molbap · 2026-01-12T14:27:16Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,


seems that both of these are unused no?

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

molbap · 2026-01-12T14:32:11Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        # 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)
+            sources.append(self.decoder_input_proj[_len_sources](encoder_outputs[0])[-1])


I'm a bit confused here, will that run? isn't encoder_outputs[0] a list?

This is from the RT-DETR code. I think there’s a typo here. It’s currently written as (encoder_outputs[0])[-1], but I believe it should be (encoder_outputs[0][-1]). Our current configuration doesn’t hit this branch, but it seemed like a good fix to make, so I’ve made the change.

I see... due to this model and v2 having a lot in common with RT-DETR @yonigozlan is overhauling it over there #41549 if you want to take a look!

Cool, it seems that at least output_hidden_states and output_attentions can be removed. Once #41549 is merged, PP-DocLayoutV2 and PP-DocLayoutV3 will follow up with these changes.

molbap · 2026-01-12T14:35:18Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        cxcy, wh = torch.split(boxes, 2, dim=-1)
+        boxes = torch.cat([cxcy - 0.5 * wh, cxcy + 0.5 * wh], dim=-1)


full words for variable names, please

molbap · 2026-01-12T14:35:44Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        B, N, _ = inputs.shape
+        proj = self.dense(inputs).reshape([B, N, 2, self.head_size])
+        proj = self.dropout(proj)
+        qw, kw = proj[..., 0, :], proj[..., 1, :]


same, no single letter variables

molbap · 2026-01-12T14:36:25Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        proj = self.dropout(proj)
+        qw, kw = proj[..., 0, :], proj[..., 1, :]
+
+        logits = torch.einsum("bmd,bnd->bmn", qw, kw) / (self.head_size**0.5)  # [B, N, N]


In 98% of situations we don't want einsum paths in inference code if an alternative exists, for both readability and performance reasons

molbap · 2026-01-12T15:12:15Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+    def forward(self, x):
+        # use 'x * F.sigmoid(x)' replace 'silu'
+        x = self.bn(self.conv(x))
+        y = x * F.sigmoid(x)
+        return y


The standard interface looks more like this (example from RTDetr which you can use directly in modular, actually):

class RTDetrResNetConvLayer(nn.Module): def __init__( self, in_channels: int, out_channels: int, kernel_size: int = 3, stride: int = 1, activation: str = "relu" ): super().__init__() self.convolution = nn.Conv2d( in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, bias=False ) self.normalization = nn.BatchNorm2d(out_channels) self.activation = ACT2FN[activation] if activation is not None else nn.Identity() def forward(self, input: Tensor) -> Tensor: hidden_state = self.convolution(input) hidden_state = self.normalization(hidden_state) hidden_state = self.activation(hidden_state) return hidden_state

and you can tune the activation function you want (sigmoid here) based on config.

molbap · 2026-01-12T15:57:38Z

src/transformers/models/pp_doclayout_v3/image_processing_pp_doclayout_v3_fast.py

+def get_order_seqs(order_logits):
+    order_scores = torch.sigmoid(order_logits)
+    batch_size, sequence_length, _ = order_scores.shape
+
+    one = torch.ones((sequence_length, sequence_length), dtype=order_scores.dtype, device=order_scores.device)
+    upper = torch.triu(one, 1)
+    lower = torch.tril(one, -1)
+
+    Q = order_scores * upper + (1.0 - order_scores.transpose(1, 2)) * lower
+    order_votes = Q.sum(dim=1)
+
+    order_pointers = torch.argsort(order_votes, dim=1)
+    order_seq = torch.full_like(order_pointers, -1)
+    batch = torch.arange(batch_size, device=order_pointers.device)[:, None]
+    order_seq[batch, order_pointers] = torch.arange(sequence_length, device=order_pointers.device)[None, :]
+
+    return order_seq


a few things here that got me thinking.( this should be likely a private method). instead of transposition you could use tril.
Then if I understand this correctly, I don't think we need to materialize Q, you could just do something like this?

order_votes = ( order_scores.triu(diagonal=1).sum(dim=1) + (1.0 - order_scores.transpose(1, 2)).tril(diagonal=-1).sum(dim=1) )

and the permutation after argsort can be something like

order_seq = torch.empty_like(order_pointers) ranks = torch.arange(sequence_length, device=order_pointers.device, dtype=order_pointers.dtype).expand(batch_size, -1) order_seq.scatter_(1, order_pointers, ranks)

excellent suggestion, thank you. I’ve made the change.

zhang-prog · 2026-01-13T12:37:30Z

@molbap
Thank you for your effort! I’ve made changes based on your comments.
PTAL.

molbap

Added another review, now let's see with the DETR refactor changes that will be merged very soon!

molbap · 2026-01-13T13:00:27Z

docs/source/en/model_doc/pp_doclayout_v3.md

+
+## Overview
+
+TBD.


Will need a small abstract, and a code usage snippet

Sure, but that may be later, because the model has not been released yet.

current usage snippets are good!

molbap · 2026-01-13T13:00:47Z

docs/source/en/model_doc/pp_doclayout_v3.md

@@ -0,0 +1,47 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.


Let's update when relevant

Suggested change

<!--Copyright 2025 The HuggingFace Team. All rights reserved.

<!--Copyright 2026 The HuggingFace Team. All rights reserved.

molbap · 2026-01-13T13:21:45Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        self.resample = resample
+
+    def _get_order_seqs(self, order_logits):
+        order_scores = torch.sigmoid(order_logits)


much better. Let's add a small docstring to explain what that does for later code inspectors. However I just noticed that the ImageProcessor (not Fast) was using torch + torchvision, the idea is that the fast image processor use torch/torchvision and is as GPU-compatible as possible, but the "slow" image processor is cpu-bound, using PIL and numpy operations only.

So the Fast Image processor (which should be the default especially if it matches the processing) is ok, but if you want to include a "slow" one, it should be without torch/torchvision

emmm. order_logits is a tensor, so, can I refer to RT-DETR and use requires_backends(self, ["torch"]) to solve this problem?

ah yes indeed, you do need to do some torch operations. In this case, I suggest simply not having the non-fast processor, it doesn't make much sense to have it here. In the mapping, it will be (None, ...ImageProcessorFast)

well, so, should I remove PPDocLayoutV3ImageProcessor and only keep PPDocLayoutV3ImageProcessorFast?

See my comment in auto, if there is no strict reason to keep a slow processor, we can IMO remove the slow one

molbap · 2026-01-13T13:27:51Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+
+        logits = (queries @ keys.transpose(-2, -1)) / (self.head_size**0.5)
+        lower = torch.tril(torch.ones([sequence_length, sequence_length], dtype=logits.dtype, device=logits.device))
+        logits = logits - lower.unsqueeze(0) * 1e4


is this 1e4 value hardcoded? would be nice to name it and have it in configuration

We’ll apply sigmoid later, so the upper triangular matrix should be masked with a very small value, no additional configuration needed.

molbap · 2026-01-13T13:50:09Z

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

+            for key, tensor in inputs_dict.items():
+                if tensor.dtype == torch.float32:
+                    inputs_dict[key] = tensor.to(dtype)


why are we converting to the wanted dtype only in float32?

molbap · 2026-01-13T13:58:03Z

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

+# TODO:
+# @require_torch
+# @require_vision
+# @slow
+# class PPDocLayoutV3ModelIntegrationTest(unittest.TestCase):


TODO I suppose! The best option is to take an image and output you expect, you can put the image on a hub dataset and request it in the test, many tests are build like that - it's the most reliable way

same as model_doc

molbap · 2026-01-13T13:58:16Z

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

@@ -0,0 +1,446 @@
+# coding = utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.


Suggested change

# Copyright 2025 The HuggingFace Inc. team. All rights reserved.

# Copyright 2026 The HuggingFace Inc. team. All rights reserved.

molbap · 2026-01-13T13:58:29Z

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

+if is_vision_available():
+    pass


Suggested change

if is_vision_available():

pass

molbap · 2026-01-13T14:27:06Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        return logits
+
+
+class PPDocLayoutV3MultiscaleDeformableAttention(RTDetrMultiscaleDeformableAttention):


could this inherit from rt detr v2 rather than v1?

What is the reason for this change? this MultiscaleDeformableAttention inheriting from rt_detr_v2 is not working, cause our model’s structure is like rt_detr rather than rt_detr_v2.

it was just out of curiosity as v2 is a bit more "modern". In any case for now it seems good, we will need #41549 to be merged and then we can merge this !

as explained, in the end no need for merging this

zhang-prog · 2026-01-14T10:53:02Z

@molbap
Model overview and PPDocLayoutV3ModelIntegrationTest cannot be completed now, but they are coming soon.
PTAL.

vasqu

We really need the rt detr refactor to make this aligned with our modern code base cc @yonigozlan to keep this model in mind

In general, I've added a lot of comments to what ideally it should be. I know that some stuff cannot be changed due to how we depend on rt detr

vasqu · 2026-01-20T13:27:29Z

docs/source/en/model_doc/pp_doclayout_v3.md

+rendered properly in your Markdown viewer.
+
+-->
+*This model was released on 2026-x-x and added to Hugging Face Transformers on 2026-x-x.*


Just fyi, we changed some CI rules so might complain, you can add arbitrary values first if needed (for release)

Okay, let me fill it in first. I’ll come back to modify it later when merging.

src/transformers/models/auto/image_processing_auto.py

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

vasqu · 2026-01-20T13:52:38Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        self.resample = resample
+
+    def _get_order_seqs(self, order_logits):
+        order_scores = torch.sigmoid(order_logits)


See my comment in auto, if there is no strict reason to keep a slow processor, we can IMO remove the slow one

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

zhang-prog · 2026-01-21T14:39:47Z

@vasqu I have modified the code according to your suggestions. PTAL.

molbap

Hey just to clarify @zhang-prog, after internal talking with @vasqu and @yonigozlan - we will merge as-is in the end, not waiting for the DETRE refactor to be done, that way code can be up as soon as possible and we'll refactor later. so once all comments are addressed it should be fine. Let's make sure it's as polished as possible before the finish line!

molbap · 2026-01-16T09:06:02Z

docs/source/en/model_doc/pp_doclayout_v3.md

+
+## Overview
+
+TBD.


current usage snippets are good!

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

molbap · 2026-01-21T14:18:40Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        return logits
+
+
+class PPDocLayoutV3MultiscaleDeformableAttention(RTDetrMultiscaleDeformableAttention):


as explained, in the end no need for merging this

molbap · 2026-01-21T14:42:05Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+
+
+@auto_docstring
+class PPDocLayoutV3PreTrainedModel(PreTrainedModel):


suggestion on that: why not RTDetrPreTrainedModel? from my analysis it is close enough, just one initialization that might be needed to override

molbap · 2026-01-21T14:45:45Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+                init.zeros_(module.weight.data[module.padding_idx])
+
+
+def mask_to_box_coordinate(mask, dtype):


for later (cc @yonigozlan ) I think we can extract a shared utility to image_utils around methods like this one

vasqu

Happy with the current state, mainly nits left re abbreviations and possibly splitting the config if bandwidth allows us to

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

vasqu · 2026-01-21T15:21:37Z

tests/models/pp_doclayout_v3/test_modeling_pp_doclayout_v3.py

@@ -0,0 +1,446 @@
+# coding = utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.


src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

vasqu · 2026-01-21T15:40:12Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+
+@dataclass
+@auto_docstring
+class PPDocLayoutV3ForObjectDetectionOutput(ModelOutput):


Could we inherit from something? Also seeing some of the output types inheriting already - maybe we can reduce repeating some parts then, not sure if autodoc handles this properly cc @yonigozlan

zhang-prog · 2026-01-22T10:08:25Z

@molbap @vasqu I’ve updated the code based on your suggestions. Let’s take a look at the latest version!

HuggingFaceDocBuilderDev · 2026-01-22T12:45:24Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

molbap

Good for me too! two nits related to model name in the doc, and an import thing

src/transformers/models/pp_doclayout_v3/__init__.py

docs/source/en/model_doc/pp_doclayout_v3.md

github-actions · 2026-01-22T13:12:44Z

CI Results

Workflow Run ⚙️

✅ No failing test specific to this PR 🎉 !

vasqu · 2026-01-22T13:14:50Z

@zhang-prog Are we waiting for the weights or will they be released after merge? Re slow integration tests

zhang-prog · 2026-01-22T13:22:37Z

@zhang-prog Are we waiting for the weights or will they be released after merge? Re slow integration tests

@vasqu We will get back to you after we have an internal discussion.

vasqu · 2026-01-22T13:23:25Z

No worries, you can also write us on slack then 🤗

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

zhang-prog · 2026-01-27T14:39:03Z

@molbap hi, Pablo, we have made three changes, and this should be the final update:

We overrode the _preprocess method, because we require self.resize(..., antialias=False) to approximate the behavior of cv2.resize in order to ensure model accuracy, but the antialias argument cannot be passed in the current implementation.
The decoder_order_head has been changed from a single nn.Linear layer to a stack of nn.Linear layers, with the depth defined by config.decoder_layers.
We added polygon_points to the post-processing stage, which are derived from the model’s output masks.

Ready for review, PTAL!

github-actions · 2026-01-27T14:47:40Z

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=43098&sha=cbbc37

molbap

Mostly naming-related comments + an important thing, prefixes to be fixed. Thank you!

molbap · 2026-01-27T15:05:07Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(


@zhang-prog if you have time 🙏

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

molbap · 2026-01-27T15:46:05Z

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

+    _no_split_modules = [r"PPDocLayoutV3HybridEncoder", r"PPDocLayoutV3DecoderLayer"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):


bumping 👀

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

vasqu

A lot of smaller things, 1 thing I'm unsure about is whether the model should tie weights - at least it looks like that to me

src/transformers/models/pp_doclayout_v3/__init__.py

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py

github-actions · 2026-01-29T09:40:25Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, pp_doclayout_v3

molbap · 2026-01-29T09:45:10Z

run-slow: pp_doclayout_v3

github-actions · 2026-01-29T09:46:16Z

This comment contains run-slow, running the specified jobs:

models: ["models/pp_doclayout_v3"]
quantizations: []

github-actions · 2026-01-29T09:55:09Z

CI Results

Workflow Run ⚙️

✅ No failing test specific to this PR 🎉 !

vasqu · 2026-01-29T09:56:04Z

Merging, CI is on fire these days and it only touches the new model which passes all tests - thanks a lot @zhang-prog

zhang-prog added 2 commits January 4, 2026 20:09

init

9b8536a

add tests

358db84

ArthurZucker reviewed Jan 5, 2026

View reviewed changes

update

4f19273

zhang-prog requested a review from ArthurZucker January 7, 2026 08:42

yonigozlan added the New model label Jan 7, 2026

molbap self-assigned this Jan 12, 2026

molbap reviewed Jan 12, 2026

View reviewed changes

zhang-prog added 2 commits January 13, 2026 20:33

update

43d1cbf

add model_doc

1eef91a

zhang-prog requested a review from molbap January 13, 2026 12:37

molbap reviewed Jan 13, 2026

View reviewed changes

update

8da28db

zhang-prog requested a review from molbap January 14, 2026 10:53

vasqu reviewed Jan 20, 2026

View reviewed changes

update

2ef0f5b

molbap reviewed Jan 21, 2026

View reviewed changes

vasqu approved these changes Jan 21, 2026

View reviewed changes

zhang-prog added 2 commits January 22, 2026 18:04

update

dbfc8ed

Merge remote-tracking branch 'origin/main' into feat/pp_doclayout_v3

64e97a2

zhang-prog and others added 4 commits January 22, 2026 18:20

Update to adapt to the latest check_code_quality.

036d478

Set to use

0d6c65d

mark as xfail

20ad456

fix wrong doc link

cd97341

molbap approved these changes Jan 22, 2026

View reviewed changes

src/transformers/models/pp_doclayout_v3/__init__.py Outdated Show resolved Hide resolved

docs/source/en/model_doc/pp_doclayout_v3.md Outdated Show resolved Hide resolved

docs/source/en/model_doc/pp_doclayout_v3.md Outdated Show resolved Hide resolved

fix nits

6cd1b46

remove padding_left

7f25cd5

molbap reviewed Jan 26, 2026

View reviewed changes

src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py Outdated Show resolved Hide resolved

molbap mentioned this pull request Jan 26, 2026

[Model] Add PP-DocLayoutV2 Model Support #43018

Open

update

cbbc379

molbap reviewed Jan 27, 2026

View reviewed changes

zhang-prog added 2 commits January 28, 2026 00:59

update

bc2b9bd

Merge remote-tracking branch 'origin/main' into feat/pp_doclayout_v3

7595ba8

vasqu reviewed Jan 27, 2026

View reviewed changes

zhang-prog added 6 commits January 28, 2026 01:24

use is_cv2_available

9f95f83

fix nits

3de4614

move _tied_weights_keys

5b9046b

update

1d5bfab

Merge remote-tracking branch 'origin/main' into feat/pp_doclayout_v3

bc9061a

update

4f4f99f

zhang-prog requested review from molbap and vasqu January 29, 2026 09:44

vasqu merged commit d196ee9 into huggingface:main Jan 29, 2026
18 of 24 checks passed

		inputs_embeds: Optional[torch.FloatTensor] = None,
		decoder_inputs_embeds: Optional[torch.FloatTensor] = None,

		cxcy, wh = torch.split(boxes, 2, dim=-1)
		boxes = torch.cat([cxcy - 0.5 * wh, cxcy + 0.5 * wh], dim=-1)

		@@ -0,0 +1,47 @@
		<!--Copyright 2025 The HuggingFace Team. All rights reserved.

	<!--Copyright 2025 The HuggingFace Team. All rights reserved.
	<!--Copyright 2026 The HuggingFace Team. All rights reserved.

		@@ -0,0 +1,446 @@
		# coding = utf-8
		# Copyright 2025 The HuggingFace Inc. team. All rights reserved.

	# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
	# Copyright 2026 The HuggingFace Inc. team. All rights reserved.

		return logits


		class PPDocLayoutV3MultiscaleDeformableAttention(RTDetrMultiscaleDeformableAttention):



		@auto_docstring
		class PPDocLayoutV3PreTrainedModel(PreTrainedModel):

		init.zeros_(module.weight.data[module.padding_idx])


		def mask_to_box_coordinate(mask, dtype):


		## Overview

		TBD.


		## Overview

		TBD.

Conversation

zhang-prog commented Jan 4, 2026

Uh oh!

ArthurZucker left a comment

Choose a reason for hiding this comment

Uh oh!

zhang-prog commented Jan 7, 2026

Uh oh!

molbap left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhang-prog Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhang-prog commented Jan 13, 2026

Uh oh!

molbap left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

zhang-prog Jan 13, 2026 •

edited

Loading