model: Add DEIMv2 to Transformers by harshaljanjani · Pull Request #44339 · huggingface/transformers

harshaljanjani · 2026-02-27T18:08:53Z

What does this PR do?

→ This PR adds DEIMv2 to Transformers!
→ IMP: I've linked two notebooks: a Colab notebook here demonstrating the functionality, predictions, checkpoint conversion, and the passing test suite for all types of presets, and another notebook here that demos the detailed fine-tuning and predictions on this preset: https://huggingface.co/Intellindust/DEIMv2_HGNetv2_N_COCO.
References:
→ Model Checkpoints
→ GitHub Pages
→ GitHub Repository
→ Paper
→ All variants, including the base variant, as well as those requiring LiteEncoder and DinoV3+STA support, are now complete and working, as demonstrated in the Colab notebook (I've used one preset from each of the three types for demo) 🥳

Closes #41211 and completes #41327.

Before submitting

This PR adds a new model to Transformers.
Did you read the contributor guidelines, specifically the Pull Request section?
Was this discussed or approved via a GitHub issue or the forum? If so, please add a link.
Did you make sure to update the documentation with your changes? Here are the documentation guidelines, and here are tips on formatting docstrings.
Did you add any necessary tests?

harshaljanjani · 2026-03-01T08:25:31Z

Happy to make changes until it's production-ready; the test suite is passing locally, and all the functionality has been tested (I've left notebooks to demo the same as mentioned in the PR description). The way I pushed the PR is as follows: the changes for the base preset are included in eaef822, and the changes for the ornamentations on top of it (HGNetV2+HybridEncoder (N), HGNetV2+LiteEncoder (Pico/Femto/Atto), and DINOv3+HybridEncoder (S/M/L/X)) are included in 85c7356 :)

harshaljanjani · 2026-03-04T06:20:56Z

Just a gentle ping on this :)

Rocketknight1 · 2026-03-05T14:07:43Z

Yes, sorry! cc @NielsRogge for review, maybe @molbap

yonigozlan · 2026-03-06T02:21:05Z

Thanks for working on this @harshaljanjani ! I'll have a look in the coming days

harshaljanjani · 2026-03-10T09:07:24Z

Thanks for working on this @harshaljanjani ! I'll have a look in the coming days

Thanks @yonigozlan! I'll await your feedback 🤗

harshaljanjani · 2026-03-16T07:45:15Z

@yonigozlan @NielsRogge Just a gentle ping :)

yonigozlan · 2026-03-16T14:06:00Z

@harshaljanjani Having a look now :), thank you for your patience!

yonigozlan

Hey @harshaljanjani ! Thanks a lot for the huge work on this PR! I have only looked at the modular file for now, but there are a few things to fix there before I get to the rest. The main point is that we shouldn't need to redefine a new dinov3 based backbone, we should be able to use the existing one through load_backbone.
Also, if you could add an integration tests for each model variants in the modeling test file, that will be great to make sure that we don't break anything during the review process.

yonigozlan · 2026-03-16T20:58:18Z

+        use_spatial_tuning_adapter (`bool`, *optional*, defaults to `False`):
+            Whether to use the Spatial Tuning Adapter (STA) for DINOv2 backbone variants.


Is this always True? let's remove it if so

It's True for DINOv3 and False for HGNetV2, but this gave me a good direction to place it in Deimv2DINOv3ConvEncoder (created in the refactor) unconditionally instead; we don't need the flag :)

yonigozlan · 2026-03-16T20:59:39Z

+        dinov3_backbone_config (`dict`, *optional*):
+            Configuration dictionary for the DINOv3 ViT backbone. Passed as kwargs to `DINOv3ViTConfig`.
+        dinov3_interaction_indexes (`list[int]`, *optional*):
+            Layer indices in the DINOv3 ViT backbone from which to extract intermediate features.
+        dinov3_hidden_dim (`int`, *optional*):
+            Hidden dimension for the DINOv3 backbone projection convolutions. If `None`, uses `hidden_size` from
+            the DINOv3 ViT config.
+        dinov3_apply_layernorm (`bool`, *optional*, defaults to `False`):
+            Whether to apply LayerNorm to intermediate features extracted from the DINOv3 ViT backbone.


This should be in backbone_config, we should have one way to load the different possible backbones, not separate paths for each

Post-refactor, backbone_config is now a DINOv3ViTConfig object directly.

yonigozlan · 2026-03-16T21:00:10Z

+        backbone_type (`str`, *optional*, defaults to `"hgnetv2"`):
+            Type of backbone to use. `"hgnetv2"` uses HGNetV2, `"dinov3"` uses DINOv3 ViT backbone with STA.


This is redundant with backbone config

yonigozlan · 2026-03-16T21:04:26Z

+class Deimv2RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.scale = nn.Parameter(torch.ones(dim))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.float()
+        hidden_states = hidden_states * torch.rsqrt(hidden_states.pow(2).mean(-1, keepdim=True) + self.eps)
+        return (hidden_states * self.scale).to(input_dtype)


Let's use an existing one from the library like LlamaRMSNorm, and make the necessary weight conversion. Also add the @use_kernel_forward_from_hub("RMSNorm") decorator

Post-refactor Deimv2RMSNorm inherits from LlamaRMSNorm directly and LlamaRMSNorm already has the decorator which should be inherited as well :)

yonigozlan · 2026-03-16T22:17:55Z

+        if config.use_spatial_tuning_adapter and config.backbone_type != "dinov3":
+            self.spatial_tuning_adapter = Deimv2SpatialTuningAdapter(config)


This is weird, let's move this to the Deimv2ConvEncoder as well (if I understood correctly)

Described in the first reply regarding the flag, STA is now inside Deimv2DINOv3ConvEncoder

yonigozlan · 2026-03-16T22:19:30Z

+            if self.config.backbone_type == "dinov3":
+                proj_feats = self.dinov3_backbone(pixel_values)
+            elif self.config.encoder_type == "lite":
+                features = self.backbone(pixel_values, pixel_mask)
+                proj_feats = [source for source, mask in features]
+            else:
+                features = self.backbone(pixel_values, pixel_mask)
+                proj_feats = [self.encoder_input_proj[level](source) for level, (source, mask) in enumerate(features)]


Here we should be able to have a single backbone call:

Suggested change

if self.config.backbone_type == "dinov3":

proj_feats = self.dinov3_backbone(pixel_values)

elif self.config.encoder_type == "lite":

features = self.backbone(pixel_values, pixel_mask)

proj_feats = [source for source, mask in features]

else:

features = self.backbone(pixel_values, pixel_mask)

proj_feats = [self.encoder_input_proj[level](source) for level, (source, mask) in enumerate(features)]

proj_feats = self.backbone(pixel_values)

yonigozlan · 2026-03-16T22:20:17Z

+        level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
+
+        if self.training and self.config.num_denoising > 0 and labels is not None:
+            from ..d_fine.modeling_d_fine import get_contrastive_denoising_training_group


Let's import it at the top, and it will be copied to the modeling file thanks to modular

yonigozlan · 2026-03-16T22:20:56Z

+
+        init_reference_points = reference_points_unact.detach()
+
+        from ..d_fine.modeling_d_fine import DFineModelOutput


Let's define a Deimv2ModelOutput inheriting from DFineModelOutput above

yonigozlan · 2026-03-16T22:22:04Z

+    @property
+    def _tied_weights_keys(self):
+        keys = {
+            r"class_embed.(?![0])\d+": r"^class_embed.0",
+            "class_embed": "model.decoder.class_embed",
+            "bbox_embed": "model.decoder.bbox_embed",
+        }
+        if getattr(self.config, "share_bbox_head", False):
+            keys[r"model\.decoder\.bbox_embed\.(?![0])\d+"] = r"model.decoder.bbox_embed.0"
+            keys[r"bbox_embed.(?![0])\d+"] = r"bbox_embed.0"
+        return keys


Not sure why we would have this but let's replace with a standard:

Suggested change

@property

def _tied_weights_keys(self):

keys = {

r"class_embed.(?![0])\d+": r"^class_embed.0",

"class_embed": "model.decoder.class_embed",

"bbox_embed": "model.decoder.bbox_embed",

}

if getattr(self.config, "share_bbox_head", False):

keys[r"model\.decoder\.bbox_embed\.(?![0])\d+"] = r"model.decoder.bbox_embed.0"

keys[r"bbox_embed.(?![0])\d+"] = r"bbox_embed.0"

return keys

_tied_weights_keys = {

r"class_embed.(?![0])\d+": r"^class_embed.0",

"class_embed": "model.decoder.class_embed",

"bbox_embed": "model.decoder.bbox_embed",

}

Actually I didn't make this change because share_bbox_head varies across presets; "Pico" / "Femto" / "Atto" share bbox heads but the rest don't; happy to update this if you think the above approach would be better and I'm missing something!

Oh I see. Not sure if this would cause issue tbh but it might be ok. I'll a core maintainer see if that can work. Otherwise the alternative would be to just copy the shared bbox embeds when converting the model, but not ideal

Should be fine tbh, we rarely have these cases but when they appear it's usually modified in the init. I actually like this version more as property - although not sure why we have to use getattr as only nit.

cc @Cyrilvallez in any case if you disagree

harshaljanjani · 2026-03-17T14:47:23Z

Thanks for your time @yonigozlan; and for the amazing review round; I've addressed the comments and ensured the test suite is passing after the update 🤗❤️
I've pushed the changes for ease of review within one single commit; this includes everything updated in response to the feedback. I added replies to your comments where I'd love for you to see how I thought through things so you can provide feedback if needed :)
Hope you find this to be a solid improvement over the last!

harshaljanjani · 2026-03-17T15:08:15Z

Also, if you could add an integration tests for each model variants in the modeling test file, that will be great to make sure that we don't break anything during the review process.

Just a quick note, the test harness should already cover the different variants :)

yonigozlan

Thanks a lot for iterating on this so quickly @harshaljanjani ! This looks much better, almost good to merge on my side,
Once you have made the last changes I suggested, You can directly ping a core maintainer for a final review.
Also don't forget to merge/rebase with your next modifications!

yonigozlan · 2026-03-18T15:12:22Z

+class Deimv2SwiGLUFFN(nn.Module):
+    def __init__(self, in_features: int, hidden_features: int, out_features: int):
+        super().__init__()
+        self.w12 = nn.Linear(in_features, 2 * hidden_features)
+        self.w3 = nn.Linear(hidden_features, out_features)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        x12 = self.w12(hidden_states)
+        x1, x2 = x12.chunk(2, dim=-1)
+        hidden = F.silu(x1) * x2
+        return self.w3(hidden)


Sorry for the back and forth on this, but if we can't directly inherit from DINOv3ViTGatedMLP here anyway, let's use the format/naming of the dinov2 swigluffn. Not splitting the linear layers might result in slightly better performance too.

yonigozlan · 2026-03-18T15:48:24Z

+class Deimv2GAPFusion(nn.Module):
+    def __init__(self, config: Deimv2Config, channels: int):
+        super().__init__()
+        self.cv = Deimv2ConvNormLayer(config, channels, channels, 1, 1, activation=config.activation_function)


Can we find a better/more descriptive name than cv?

yonigozlan · 2026-03-18T16:05:31Z

+            if pixel_mask is None:
+                pixel_mask = torch.ones(((batch_size, height, width)), device=device)
+


Just a heads up that pixel_mask is not used here. This is not on you as pixel mask is not supported in dinov3 for now. I'll add it later but this is out of scope for this PR. Would you mind adding a Todo comment to pass the mask to the backbone once dinov3 supports it? thanks!

yonigozlan · 2026-03-18T16:16:34Z

+class Deimv2MLPPredictionHead(DFineMLP):
+    pass


This is not used at all

My bad, removed!

yonigozlan · 2026-03-18T16:37:48Z

+        if is_dinov3:
+            self.backbone = Deimv2DINOv3ConvEncoder(config)
+        else:
+            self.backbone = Deimv2ConvEncoder(config)


Here I think it would make more sense to name these self.conv_encoder. I know in DFine this is named backbone as well, but here we include the feature projections and sta etc., so this is more than a backbone. + it will avoid having "backbone.backbone" which I don't like in DFine

Agreed; thought it was a bit weird too. Made the broader self.conv_encoder change as and where needed :)

yonigozlan · 2026-03-18T19:33:27Z

@@ -0,0 +1,68 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.


Suggested change

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

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

yonigozlan · 2026-03-18T19:34:29Z

@@ -0,0 +1,29 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.


Suggested change

# Copyright 2025 The HuggingFace Team. All rights reserved.

# Copyright 2026 The HuggingFace Team. All rights reserved.

yonigozlan · 2026-03-18T19:34:58Z

@@ -0,0 +1,784 @@
+# Copyright 2025 The HuggingFace Inc. team.


Suggested change

# Copyright 2025 The HuggingFace Inc. team.

# Copyright 2026 The HuggingFace Inc. team.

yonigozlan · 2026-03-18T19:35:08Z

@@ -0,0 +1,1142 @@
+# 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.

yonigozlan · 2026-03-18T19:35:30Z

+# coding = utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.


Suggested change

# coding = utf-8

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

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

harshaljanjani · 2026-03-19T12:08:48Z

Thanks a lot for the review cycle! Addressed all the newer comments in the commit 4ad0dc5 and verified that there are no regressions in the test suite 🤗
Pinging @ArthurZucker for the final review as suggested above; please let me know if I've missed any additional pings :)

HuggingFaceDocBuilderDev · 2026-03-23T22:02:27Z

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.

harshaljanjani · 2026-03-24T10:00:07Z

Good day @ArthurZucker; just a gentle ping to check if there are any updates regarding the final review, thank you!
cc: @yonigozlan

yonigozlan · 2026-03-24T13:42:36Z

Also pinging @vasqu @Cyrilvallez for core maintainer review :)

harshaljanjani · 2026-04-23T15:03:21Z

Hi, could you please check this PR #45601 regarding a bug in the D-FINE loss, I think it is also relevant for DEIMv2

@m-matthias Thanks for flagging this in the PR; it turns out DEIMv2 suffered from the same coupling in the copy-over. I've pushed the fix in fb1f387: it's independent of the D-FINE fix (also cc: @Abineshabee!). You both are awesome 🤗
@vasqu Would be grateful if you could cross-check this as well whenever you have a moment, thank you :)

vasqu · 2026-04-23T15:07:36Z

SGTM, I reviewed the other PR but would like to wait on Yoni before merging. Can you sync with main here btw, we updated the linter so the false positive shouldn't appear anymore

github-actions · 2026-04-23T15:11:46Z

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

run-slow: auto, d_fine, deimv2

harshaljanjani · 2026-04-23T17:13:34Z

Can you sync with main here btw, we updated the linter so the false positive shouldn't appear anymore

Yepp; false positive's gone, awesome!

harshaljanjani · 2026-04-27T09:20:40Z

Good day; just checking in to see if there are any updates!

vasqu · 2026-04-27T13:02:01Z

run-slow: d_fine, deimv2

vasqu · 2026-04-27T13:02:39Z

@harshaljanjani sorry, no everything should be ready; one last sanity check then merging 🤗 thanks a lot for the contribution, definitely wasn't an easy model

github-actions · 2026-04-27T13:02:48Z

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

run-slow: auto, d_fine, deimv2

github-actions · 2026-04-27T13:03:22Z

Workflow Run ⚙️

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

models: ["models/d_fine", "models/deimv2"]
quantizations: []

github-actions · 2026-04-27T13:19:18Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	f7039bb7	workflow commit (merge commit)
PR	a3d9ac56	branch commit (from PR)
main	e651c68e	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

github-actions · 2026-04-27T13:21:26Z

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

run-slow: auto, d_fine, deimv2

harshaljanjani · 2026-04-27T14:25:35Z

@harshaljanjani sorry, no everything should be ready; one last sanity check then merging 🤗 thanks a lot for the contribution, definitely wasn't an easy model

Had tons of fun; thank you as well for the timely review cycles on the PR, @vasqu and @yonigozlan! 😊

* init: Add files (v1) * fix: Fix ci/circleci: check_repository_consistency * feat: Add support and test harness for all variants * fix: Fix ci/circleci: check_repository_consistency * refactor: Resolve review comments * refactor: Resolve second review round * nit: Fix copyright year * refactor: Resolve third review round * revert: Adhere to the pattern from yonigozlan * nit: Clarify the docstring * refactor: Resolve fourth review round * refactor: Closing in on the final set of nits * fix: Resolve merge conflicts * fix: Add loss override and address nits * nits: Fix minor issues * fixup their init weights * fix: Fix loss coupling issue * fix date --------- Co-authored-by: vasqu <antonprogamer@gmail.com> Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>

harshaljanjani added 4 commits February 27, 2026 22:06

init: Add files (v1)

eaef822

fix: Fix ci/circleci: check_repository_consistency

ddc1bd7

feat: Add support and test harness for all variants

85c7356

fix: Fix ci/circleci: check_repository_consistency

adc4079

harshaljanjani marked this pull request as ready for review March 1, 2026 08:26

github-actions Bot requested review from ArthurZucker and Rocketknight1 March 1, 2026 08:26

Merge branch 'main' into add-deimv2

81a3d06

yonigozlan reviewed Mar 16, 2026

View reviewed changes

refactor: Resolve review comments

39d300e

harshaljanjani requested a review from yonigozlan March 17, 2026 14:53

yonigozlan reviewed Mar 18, 2026

View reviewed changes

harshaljanjani added 2 commits March 19, 2026 12:14

Merge branch 'main' into add-deimv2

476d69f

refactor: Resolve second review round

4ad0dc5

harshaljanjani added 3 commits March 19, 2026 16:24

nit: Fix copyright year

16f2d07

Merge branch 'main' into add-deimv2

78eaf93

Merge branch 'main' into add-deimv2

dbe577b

Merge branch 'main' into add-deimv2

3629f13

evalstate mentioned this pull request Apr 23, 2026

[Auto] Add DEIMv2 to Transformers (cluster-41211-3): merged 1 of 3 PRs evalstate/transformers#11

Closed

This was referenced Apr 23, 2026

[Auto] Add DEIMv2 (cluster-41211-3): merged 1 of 2 PRs evalstate/transformers#18

Closed

[Auto] Add DEIMv2 (cluster-41211-3): merged 1 of 3 PRs evalstate/transformers#25

Closed

[Auto] Add DEIMv2 (cluster-41211-3): merged 1 of 3 PRs evalstate/transformers#30

Open

This was referenced Apr 27, 2026

[Auto] Add DEIMv2 (cluster-41211-3): merged 1 of 2 PRs evalstate/transformers#34

Open

[Auto] Super merge super-merge-example: combined 6 of 10 worktrees evalstate/transformers#40

Open

Merge branch 'main' into add-deimv2

a3d9ac5

vasqu added the New model label Apr 27, 2026

fix date

b9becfe

vasqu enabled auto-merge April 27, 2026 13:20

vasqu added this pull request to the merge queue Apr 27, 2026

vasqu removed this pull request from the merge queue due to a manual request Apr 27, 2026

vasqu added this pull request to the merge queue Apr 27, 2026

Merged via the queue into huggingface:main with commit 739b46c Apr 27, 2026
28 checks passed

harshaljanjani deleted the add-deimv2 branch April 27, 2026 14:23

		use_spatial_tuning_adapter (`bool`, optional, defaults to `False`):
		Whether to use the Spatial Tuning Adapter (STA) for DINOv2 backbone variants.

		backbone_type (`str`, optional, defaults to `"hgnetv2"`):
		Type of backbone to use. `"hgnetv2"` uses HGNetV2, `"dinov3"` uses DINOv3 ViT backbone with STA.

		if config.use_spatial_tuning_adapter and config.backbone_type != "dinov3":
		self.spatial_tuning_adapter = Deimv2SpatialTuningAdapter(config)


		init_reference_points = reference_points_unact.detach()

		from ..d_fine.modeling_d_fine import DFineModelOutput

		if pixel_mask is None:
		pixel_mask = torch.ones(((batch_size, height, width)), device=device)

		@@ -0,0 +1,68 @@
		<!--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 +1,29 @@
		# 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 +1,784 @@
		# Copyright 2025 The HuggingFace Inc. team.

	# Copyright 2025 The HuggingFace Inc. team.
	# Copyright 2026 The HuggingFace Inc. team.

		@@ -0,0 +1,1142 @@
		# Copyright 2025 The HuggingFace Inc. team. All rights reserved.

		# coding = utf-8
		# Copyright 2025 The HuggingFace Inc. team. All rights reserved.

Conversation

harshaljanjani commented Feb 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Before submitting

Uh oh!

harshaljanjani commented Mar 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

harshaljanjani commented Mar 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Rocketknight1 commented Mar 5, 2026

Uh oh!

yonigozlan commented Mar 6, 2026

Uh oh!

harshaljanjani commented Mar 10, 2026

Uh oh!

harshaljanjani commented Mar 16, 2026

Uh oh!

yonigozlan commented Mar 16, 2026

Uh oh!

yonigozlan 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!

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!

harshaljanjani commented Mar 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

harshaljanjani commented Mar 17, 2026

Uh oh!

yonigozlan left a comment • 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

harshaljanjani commented Feb 27, 2026 •

edited

Loading

harshaljanjani commented Mar 1, 2026 •

edited

Loading

harshaljanjani commented Mar 4, 2026 •

edited

Loading

harshaljanjani commented Mar 17, 2026 •

edited

Loading

yonigozlan left a comment •

edited

Loading

harshaljanjani commented Mar 19, 2026 •

edited

Loading

harshaljanjani commented Apr 23, 2026 •

edited

Loading