PerceptionLM #37878

shuminghu · 2025-04-29T23:39:54Z

This PR implements PerceptionLM released by Meta:
https://github.com/facebookresearch/perception_models

github-actions · 2025-04-29T23:40:05Z

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the Ready for review button (at the bottom of the PR page). This will assign reviewers and trigger CI.

zucchini-nlp

Yay, super excited to get the model shipped! I know it is early to review, I noticed the model doesn't have a modular file yet. I recommend to use modular transformers to add the model, it will allow you to inherit from any similar model in transformers and you won't have to rewrite the whole class

Also it makes the review process easier and faster, since we see what are the main differences between PE and other existing model 😉

shuminghu · 2025-04-30T17:56:46Z

Yay, super excited to get the model shipped! I know it is early to review, I noticed the model doesn't have a modular file yet. I recommend to use modular transformers to add the model, it will allow you to inherit from any similar model in transformers and you won't have to rewrite the whole class

Also it makes the review process easier and faster, since we see what are the main differences between PE and other existing model 😉

I see. Let me take a look. The classes here were added automatically (first commit: plm template) from running this command:

transformers-cli add-new-model-like
What is the model you would like to duplicate? Please provide the lowercase `model_type` (e.g. roberta): llava
What is the name (with no special casing) for your new model in the paper (e.g. RoBERTa)? PerceptionLM
What identifier would you like to use for the `model_type` of this model?  [perceptionlm] perception_lm
What lowercase name would you like to use for the module (folder) of this model?  [perceptionlm] perception_lm
What prefix (camel-cased) would you like to use for the model classes of this model (e.g. Roberta)?  [PerceptionLM] 
What prefix (upper-cased) would you like to use for the constants relative to this model?  [PERCEPTIONLM] PERCEPTION_LM
What will be the name of the config class for this model?  [PerceptionLMConfig] 
Please give a checkpoint identifier (on the model Hub) for this new model (e.g. facebook/FacebookAI/roberta-base): facebook/Perception-LM-1B
Will your new model use the same processing class as llava (LlamaTokenizer, LlavaProcessor) (yes/no)? no
What will be the name of the tokenizer class for this model?  [PerceptionLMTokenizer] 
What will be the name of the processor class for this model?  [PerceptionLMProcessor] 
Should we add # Copied from statements when creating the new modeling file (yes/no)?  [yes] 
Should we add a version of your new model in all the frameworks implemented by llava (['pt']) (yes/no)?  [yes] 
The constants at the start of the new tokenizer file created needs to be manually fixed. If your new model has a tokenizer fast, you will also need to manually add the converter in the `SLOW_TO_FAST_CONVERTERS` constant of `convert_slow_tokenizer.py`.

zucchini-nlp · 2025-05-01T10:09:58Z

Yeah, that way is correct and usually copies an existing similar model. In this case Llava was written without modular as it was almost the first VLM in transformers :)

zucchini-nlp

Thanks for great work @shuminghu !

My main concern currently is to make the image processing and the vision tower more aligned with transformers standards. I left some comments below on how we could load the model, lmk if that works

src/transformers/models/auto/modeling_auto.py

src/transformers/models/perception_lm/modular_perception_lm.py

src/transformers/models/perception_lm/processing_perception_lm.py

test.py

zucchini-nlp · 2025-05-20T11:38:58Z

tests/models/perception_lm/test_modeling_perception_lm.py

+    Model tester for `PerceptionLMForConditionalGeneration`.
+    """
+
+    all_model_classes = (PerceptionLMForConditionalGeneration,) if is_torch_available() else ()


all_generative_model_classes as well, fast generation tests are needed to make sure we don't break anything in the future. We don't normally run integration tests for all PRs

src/transformers/models/perception_lm/configuration_perception_lm.py

src/transformers/models/perception_lm/modular_perception_lm.py

zucchini-nlp · 2025-05-26T07:33:26Z

@shuminghu ping me when it is ready for review and can you mark it PR as "ready for review' as well'

shuminghu · 2025-05-27T21:53:05Z

Yes. Absolutely. Circling back to this now.

…

On Mon, May 26, 2025 at 12:33 AM Raushan Turganbay ***@***.***> wrote: *zucchini-nlp* left a comment (huggingface/transformers#37878) <#37878 (comment)> @shuminghu <https://github.com/shuminghu> ping me when it is ready for review and can you mark it PR as "ready for review' as well' — Reply to this email directly, view it on GitHub <#37878 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWMMF4SLPGNHMYQTRWYENL3AK7V3AVCNFSM6AAAAAB4EIC6R6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSMBYHA2DAMZQGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

zucchini-nlp

Great job, super clean! ❤️ I left mostly comments about nits, like deleting unnecessary files and print statements

The only major comment to address is about moving contents from image_trasnfom.py to the processing file, and making sure it follows the transformers format style (i.e. uses torchvision.F without compose)

zucchini-nlp · 2025-06-13T09:06:37Z

docs/source/en/model_doc/perception_lm.md

+## Overview
+
+The PerceptionLM model was proposed in [<INSERT PAPER NAME HERE>](<INSERT PAPER LINK HERE>) by <INSERT AUTHORS HERE>.
+<INSERT SHORT SUMMARY HERE>


docs to be filled before merging

src/transformers/models/auto/image_processing_auto.py

src/transformers/models/auto/modeling_auto.py

zucchini-nlp · 2025-06-13T09:08:53Z

src/transformers/models/perception_lm/configuration_perception_lm.py

+        use_cls_token=True,
+        architecture="vit_pe_core_large_patch14_336",
+        width=1024,
+        img_size=[448, 448],


can't use mutables as defaults for args in python, same for ref_feat_shape. Lets make it a tuple

I changed them from Tuple coz a unit test failed

tests/models/perception_lm/test_modeling_perception_lm.py::PerceptionLMForConditionalGenerationModelTest::test_config FAILED [ 12%]

============================================================================= short test summary info ============================================================================== FAILED tests/models/perception_lm/test_modeling_perception_lm.py::PerceptionLMForConditionalGenerationModelTest::test_config - AssertionError: {'ima[1414 chars]'', 'model_type': 'perception_encoder', 'use_c[3584 chars]alse} != {'ima[1414 chars]'', 'use_cls_token': True, 'architecture': 'vi[3584 chars]... =========================================================================== 1 failed, 1 warning in 1.19s ============================

it turns out config serialized from json produces List which is not equal to original Tuple.

we can initialize te default if after as in img_size = img_size is img_size is not None else [448, 448]

src/transformers/models/perception_lm/configuration_perception_lm.py

test.py

zucchini-nlp · 2025-06-13T09:52:00Z

tests/models/perception_lm/test_processor_perception_lm.py

+    import torch
+
+
+TEST_MODEL_PATH = "shumingh/plm_1b_hf"


to update to actual ckpt after the model is released and public

zucchini-nlp · 2025-06-13T09:52:55Z

tests/models/perception_lm/test_modeling_perception_lm.py

+        # input_ids[
+        #     :, self.num_image_tokens : self.num_video_tokens + self.num_image_tokens
+        # ] = config.video_token_id
+


video_inference doesn't work? We would need tests for all modalities

we don't support mixing image and video in the same prompt

then we can add a separate Tester for video only, so it is tested

Added back. it doesnt matter here.

zucchini-nlp · 2025-06-13T09:53:14Z

tests/models/perception_lm/test_modeling_perception_lm.py

+        pass
+
+
+TEST_MODEL_PATH = "shumingh/plm_1b_hf"


same, to be updated before merge, so we dont forget

tests/models/perception_lm/test_modeling_perception_lm.py

zucchini-nlp

Thanks for iterating! Approving as the PR looks good, left only minor comments for readability. Let's make the CI green and ask for last review from core maintainers

For the failing integration tests that a timm model isn't found, cc @ydshieh . Which version on timm do we have in the runners and can we update it? We'll also have another big release soon based on timm as a backbone

docs/source/en/model_doc/perception_lm.md

src/transformers/models/auto/image_processing_auto.py

zucchini-nlp · 2025-06-17T07:29:34Z

src/transformers/models/perception_lm/configuration_perception_lm.py

+        use_cls_token=True,
+        architecture="vit_pe_core_large_patch14_336",
+        width=1024,
+        img_size=[448, 448],


we can initialize te default if after as in img_size = img_size is img_size is not None else [448, 448]

src/transformers/models/perception_lm/image_processing_perception_lm_fast.py

zucchini-nlp · 2025-06-17T07:47:12Z

src/transformers/models/perception_lm/image_processing_perception_lm_fast.py

+                    thumbnails, _ = self.resize(stacked_images, self.tile_size, max_num_tiles=1)
+                    images_for_tiling, (tiles_w, tiles_h) = self.resize(
+                        stacked_images, self.tile_size, max_num_tiles=self.max_num_tiles
+                    )


we should be able to configure self.tile_size at call time and it is guaranteed to be in the input args. So we need to expand the signature args and use tile_size. Same for max_num_tiles

ydshieh · 2025-06-17T09:07:29Z

For the failing integration tests that a timm model isn't found, cc @ydshieh . Which version on timm do we have in the runners and can we update it? We'll also have another big release soon based on timm as a backbone

Hi, it doesn't seems just integration tests but all tests, if we are talking about

RuntimeError: Unknown model (vit_pe_core_large_patch14_336)

In the runner, we have

timm==1.0.15

(this info. could be find in the job step Show installed libraries and their versions)

However, we can not use a model like vit_pe_core_large_patch14_336 in non-slow tests (for circleci for example). Usually with HF models, we set a config with very small values for some attributes, and use that config to create a very tiny model on the fly.

For integration tests, we can use it.

zucchini-nlp · 2025-06-17T09:31:13Z

Right, we need to modify non-integration tests @shuminghu . And for the integration ones, do we need to add a decorator for timm version so it doesn't flood nightlies with failures or do we update timm in runners?

shuminghu · 2025-06-17T15:41:31Z

vit_pe_core_large_patch14_336 Here is architecture template. But actual number of layers and dims are small numbers not ViT-L scale in non-integration tests.

…

On Tue, Jun 17, 2025 at 2:31 AM Raushan Turganbay ***@***.***> wrote: *zucchini-nlp* left a comment (huggingface/transformers#37878) <#37878 (comment)> Right, we need to modify non-integration tests @shuminghu <https://github.com/shuminghu> . And for the integration ones, do we need to add a decorator for timm version so it doesn't flood nightlies with failures or do we update timm in runners? — Reply to this email directly, view it on GitHub <#37878 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWMMF5VXCCHJGODU7YWDPL3D7N7NAVCNFSM6AAAAAB4EIC6R6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSNZZGY2DENRXGU> . You are receiving this because you were mentioned.Message ID: ***@***.***>

shuminghu · 2025-06-26T05:19:19Z

For the failing integration tests that a timm model isn't found, cc @ydshieh . Which version on timm do we have in the runners and can we update it? We'll also have another big release soon based on timm as a backbone

Hi, it doesn't seems just integration tests but all tests, if we are talking about

RuntimeError: Unknown model (vit_pe_core_large_patch14_336)

In the runner, we have

timm==1.0.15

(this info. could be find in the job step Show installed libraries and their versions)

However, we can not use a model like vit_pe_core_large_patch14_336 in non-slow tests (for circleci for example). Usually with HF models, we set a config with very small values for some attributes, and use that config to create a very tiny model on the fly.

For integration tests, we can use it.

@ydshieh vit_pe_core_large_patch14_336 is actually not used in non-integration test despite the naming.
With latest TimmWrapperConfig support, it is specified as following in non-integration tests, where architecture is merely a template and model dim and depth is specified as model_args. But yea, all of this requires timm's source code up-to-date to a week ago.

        vision_config={
            "architecture": "vit_pe_core_large_patch14_336",
            "model_args": {
                "embed_dim": 64,
                "img_size": (14, 14),
                "depth": 2,
                "global_pool": "",
                "use_post_transformer_norm": False,
                "init_values": 0.1,
                "ref_feat_shape": (1, 1),
            },
        },

ydshieh · 2025-06-27T13:11:12Z

Happy to upgrade timm in docker if @zucchini-nlp and @qubvel confirm it is necessary.

Does this mean we also have to pin timm version in setup.py so users will have the correct timm version?

(And this requirement is specific to this new model or we need it anyway in general?)

qubvel · 2025-06-27T13:46:11Z

We also need the latest timm for gemma3n, because mobilenetv5 was added just recently, but not sure if we should pin reqs strict to the latest version, maybe just add version check with correct error message

ydshieh · 2025-06-27T13:56:39Z

we have timm==1.0.15 at this moment, and we have v1.0.16 released 16 hours ago).

@shuminghu Are the failing tests caused by the timm==1.0.15 ...?

shuminghu · 2025-06-27T16:42:45Z

we have timm==1.0.15 at this moment, and we have v1.0.16 released 16 hours ago).

@shuminghu Are the failing tests caused by the timm==1.0.15 ...?

@ydshieh Right. v1.0.16 would be good for me for CI.

From Release v1.0.16 change log:
...
Add EVA ViT based PE (Perceptual Encoder) impl by @rwightman in huggingface/pytorch-image-models#2487

--- upate ----
Thanks @ydshieh! I just saw CI is passing. So it must have been updated.

HuggingFaceDocBuilderDev · 2025-06-30T07:27:54Z

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.

Cyrilvallez

Alright! Great great work, super clean 🤗🤗 Just a few comments to polish a bit, but this is very nice already and almost ready to be shipped!👌

Cyrilvallez · 2025-07-04T09:04:18Z

src/transformers/models/perception_lm/modular_perception_lm.py

+
+from transformers.generation.utils import GenerationMixin


Let's be consistent and use relative imports everywhere!

Cyrilvallez · 2025-07-04T09:06:31Z

src/transformers/models/perception_lm/modular_perception_lm.py

+_CONFIG_FOR_DOC = "PerceptionLMConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "facebook/Perception-LM-1B"


Can be removed as it does not seem to be used (they were here before we had auto_docstring mostly)

Cyrilvallez · 2025-07-04T09:06:51Z

src/transformers/models/perception_lm/modular_perception_lm.py

+class AdaptiveAvgPooling(nn.Module):
+    def __init__(self, pooling_ratio=2):
+        super(AdaptiveAvgPooling, self).__init__()


Suggested change

class AdaptiveAvgPooling(nn.Module):

def __init__(self, pooling_ratio=2):

super(AdaptiveAvgPooling, self).__init__()

class PerceptionLMAdaptiveAvgPooling(nn.Module):

def __init__(self, pooling_ratio=2):

super().__init__()

Cyrilvallez · 2025-07-04T09:07:54Z

src/transformers/models/perception_lm/modular_perception_lm.py

+        self.pooling = (
+            AdaptiveAvgPooling(config.projector_pooling_ratio) if config.projector_pooling_ratio > 1 else nn.Identity()
+        )


Do we have checkpoints where it's actually > 1? I see it's 1 by default in the config. If it's not used in any real checkpoint, let's remove this layer for simplicity!

Yes we use pooling ratio 2.
https://huggingface.co/facebook/Perception-LM-1B/blob/main/params.json#L19

Cyrilvallez · 2025-07-04T09:09:52Z

src/transformers/models/perception_lm/modular_perception_lm.py

+        self.projector = nn.ModuleList(
+            [
+                nn.Linear(
+                    in_features=input_size,
+                    out_features=output_size,
+                    bias=True,
+                ),
+                nn.GELU(),
+                nn.Linear(
+                    in_features=output_size,
+                    out_features=output_size,
+                    bias=True,
+                ),
+            ]
+        )


Here if possible, I'd rather we unwrap those layers and use them directly in the forward as well instead if iterating, for clarity (would require to update the converter as well)

Cyrilvallez · 2025-07-04T09:17:57Z

src/transformers/models/perception_lm/modular_perception_lm.py

+    # Make modules available throught conditional class for BC
+    @property
+    def language_model(self):
+        return self.model.language_model
+
+    @property
+    def vision_tower(self):
+        return self.model.vision_tower
+
+    @property
+    def multi_modal_projector(self):
+        return self.model.multi_modal_projector


We should not need those

Oh. I added this back (BC). This was added to fix this test

@require_torch_sdpa def test_sdpa_can_dispatch_composite_models(self): """ Tests if composite models dispatch correctly on SDPA/eager when requested so when loading the model. This tests only by looking at layer names, as usually SDPA layers are called "SDPAAttention". In contrast to the above test, this one checks if the "config._attn_implamentation" is a dict after the model is loaded, because we manually replicate requested attn implementation on each sub-config when loading. See https://github.com/huggingface/transformers/pull/32238 for more info The test tries to cover most general cases of composite models, VLMs with vision and text configs. Any model that has a different set of sub-configs has to overwrite this test. """ if not self.has_attentions: self.skipTest(reason="Model architecture does not support attentions") if not self._is_composite: self.skipTest(f"{self.all_model_classes[0].__name__} does not support SDPA") for model_class in self.all_model_classes: config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() model = model_class(config) with tempfile.TemporaryDirectory() as tmpdirname: model.save_pretrained(tmpdirname) model_sdpa = model_class.from_pretrained(tmpdirname) model_sdpa = model_sdpa.eval().to(torch_device) vision_model_names = {"visual", "image_tower", "vision_tower", "vision_model"} language_model_names = {"language_model", "model", "text_model"} > vision_model_name = [name for name in vision_model_names if hasattr(model_sdpa, name)][0] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ E IndexError: list index out of range

I mentioned test_sdpa_can_dispatch_composite_models in the comment.

Cyrilvallez · 2025-07-04T09:20:34Z

src/transformers/models/perception_lm/modular_perception_lm.py

+@auto_docstring
+class PerceptionLMModel(LlavaModel):


LlavaModel has a mapping that we should not need here, and that we should overwrite to empty explicitly

Suggested change

@auto_docstring

class PerceptionLMModel(LlavaModel):

@auto_docstring

class PerceptionLMModel(LlavaModel):

_checkpoint_conversion_mapping = {}

Cyrilvallez · 2025-07-04T09:21:19Z

src/transformers/models/perception_lm/modular_perception_lm.py

+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,


Similar as before for return_dict with the decorator! Let's remove it everywhere!

removed return_dict from all methods.

Cyrilvallez · 2025-07-04T09:23:53Z

tests/models/perception_lm/test_modeling_perception_lm.py

+    @unittest.skip("ViT PE cannot be tested with meta device")
+    def test_can_be_initialized_on_meta(self):
+        pass
+
+    @unittest.skip("ViT PE cannot be tested with meta device")
+    def test_can_load_with_meta_device_context_manager(self):
+        pass


Is that due to timm?

Yes it is. Mentioned timm in reason string.

Cyrilvallez · 2025-07-04T09:27:27Z

tests/models/perception_lm/test_modeling_perception_lm.py

+    def test_training(self):
+        self.all_model_classes = (PerceptionLMForConditionalGeneration,) if is_torch_available() else ()
+        super().test_training()
+
+    def test_training_gradient_checkpointing(self):
+        self.all_model_classes = (PerceptionLMForConditionalGeneration,) if is_torch_available() else ()
+        super().test_training_gradient_checkpointing()
+
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        self.all_model_classes = (PerceptionLMForConditionalGeneration,) if is_torch_available() else ()
+        super().test_training_gradient_checkpointing_use_reentrant()
+
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        self.all_model_classes = (PerceptionLMForConditionalGeneration,) if is_torch_available() else ()
+        super().test_training_gradient_checkpointing_use_reentrant_false()


I never remember exactly how pytest runs the tests, but I think that overriding self.all_model_classes like this may pollute some other tests using the attribute 🥲 Why do we change it? Should work with the base model as well

oh these are instantiated per test. I tested before by printing "self.all_model_classes" at the beginning of each test when running them all.

The reason is loss is not defined base model (PerceptionLMModel), only PerceptionLMForConditionalGeneration has lm_head, loss, labels defined which are used in testing. I think other models (llava_onvision...) chose to skip these tests. I kept them here as i think tests are still valid for PerceptionLMForConditionalGeneration

github-actions · 2025-07-05T05:23:41Z

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

run-slow: auto, perception_lm

github-actions bot marked this pull request as draft April 29, 2025 23:40

shuminghu changed the title ~~Perception lm~~ [WIP] Perception lm Apr 30, 2025

zucchini-nlp reviewed Apr 30, 2025

View reviewed changes

mmaaz60 mentioned this pull request May 8, 2025

Make the PLM generate multiple different responses for the same input prompt facebookresearch/perception_models#36

Open

shuminghu force-pushed the perception_lm branch from 6b0db99 to 82af49c Compare May 12, 2025 20:58

zucchini-nlp reviewed May 20, 2025

View reviewed changes

shuminghu force-pushed the perception_lm branch from d97544c to 1ffe8e5 Compare May 30, 2025 23:04

shuminghu force-pushed the perception_lm branch from c19d311 to b951470 Compare June 13, 2025 07:13

shuminghu changed the title ~~[WIP] Perception lm~~ PerceptionLM Jun 13, 2025

shuminghu marked this pull request as ready for review June 13, 2025 07:20

zucchini-nlp reviewed Jun 13, 2025

View reviewed changes

zucchini-nlp approved these changes Jun 17, 2025

View reviewed changes

qubvel mentioned this pull request Jun 17, 2025

Add kwargs for timm.create_model in TimmWrapper #38860

Merged

qubvel added the New model label Jun 17, 2025

shuminghu force-pushed the perception_lm branch from b453465 to 2fccbca Compare June 25, 2025 18:08

shuminghu force-pushed the perception_lm branch from d418f0d to 4a0f26e Compare June 27, 2025 20:45

shuminghu added 24 commits July 3, 2025 04:54

CI and doc fix

836f546

CI formatting fix

b9e5fa0

ruff fix

6b69945

ruff formatting

6c82012

ran utils/sort_auto_mappings.py

ed1dd4b

update docstring

b77f53e

more docstring udpates

eebcc7a

add vision_input_type default fallback for image processing

2c73fc4

more verbose variable naming

6ceb83a

test update

87c6ca4

Remove PE and PEConfig use AutoModel(TimmWrapper) instead

d7b47d6

Minor cleanup.

cbc1057

Minor Fix: remove any ref to PE. Ruff format and check.

d62b35d

fix docstring

f6d095a

Fix modular/model consistency.Improvex docstringfor .

508117d

Fix PerceptionLMForConditionalGenerationModelTest

bcedcc0

ruff fix

f25a2ca

fix for check_repo

e8d08e8

minor formatting

4c05fb3

dummy size arg to fix for processor test.

c74d652

Update docstring for PerceptionLMConfig

3c46d3a

Minor fixes from review feedback.

160b039

Revert some minor changes per reviewer feedback.

97fbfca

update base_model_prefix

99a1bf2

shuminghu force-pushed the perception_lm branch from d50b8c9 to 99a1bf2 Compare July 3, 2025 04:54

Cyrilvallez reviewed Jul 4, 2025

View reviewed changes

Cyrilvallez and others added 3 commits July 4, 2025 11:30

Merge branch 'main' into perception_lm

74b7995

address reviewer feedback

73c919a

fix comment in modeling file

423af52

PerceptionLM #37878

Are you sure you want to change the base?

PerceptionLM #37878

Uh oh!

Conversation

shuminghu commented Apr 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Apr 29, 2025

Uh oh!

zucchini-nlp left a comment

Choose a reason for hiding this comment

Uh oh!

shuminghu commented Apr 30, 2025

Uh oh!

zucchini-nlp commented May 1, 2025

Uh oh!

zucchini-nlp left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

zucchini-nlp commented May 26, 2025

Uh oh!

shuminghu commented May 27, 2025 via email

Uh oh!

zucchini-nlp left a comment

Choose a reason for hiding this comment

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!

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!

Uh oh!

zucchini-nlp left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ydshieh commented Jun 17, 2025

Uh oh!

zucchini-nlp commented Jun 17, 2025

Uh oh!

shuminghu commented Jun 17, 2025 via email

shuminghu commented Apr 29, 2025 •

edited

Loading

ydshieh commented Jun 27, 2025 •

edited

Loading

qubvel commented Jun 27, 2025 •

edited

Loading

shuminghu commented Jun 27, 2025 •

edited

Loading