🚨 Speed up detr image processing by guarin · Pull Request #48066 · huggingface/transformers

guarin · 2026-08-18T16:31:43Z

What does this PR do?

Speeds up preprocessing on GPU for all detr models by using group_images_by_shape. With batch size 32 and ragged inputs between 512px to 2048px I got 27-38% speed up on an H100:

model	baseline step ms	speedup-branch step ms	improvement
rtdetrv2	20.84	15.23	27% faster
rf-detr	25.85	16.37	37% faster
detr-panoptic	28.19	17.94	36% faster
conditional-detr	28.06	17.93	36% faster
dab-detr	28.77	17.97	38% faster
deformable-detr	28.16	18.01	36% faster
lw-detr	22.68	16.26	28% faster

Other minor changes:

group_images_by_shape now allows passing None as paired_input which simplifies annotation handling
remove rtdetr preprocess and inherit from detr instead (they are identical)
🚨 use antialias=False in rf-detr resize following fix that was recently added in the original rf-detr repo: https://github.com/roboflow/rf-detr/blob/5547a9360c19df7ff8cedcd29f841f7de01f2f75/src/rfdetr/detr.py#L2480-L2481

Looks like a lot of changes but only detr, rt-detr, and rf-detr changed. The rest is auto-generated.

cc @molbap @zucchini-nlp

…m/guarin/transformers into speed-up-detr-image-preprocessing

HuggingFaceDocBuilderDev · 2026-08-18T16:48:30Z

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.

guarin · 2026-08-18T17:21:57Z

run-slow: conditional_detr, deformable_detr, detr, grounding_dino, rf_detr, rt_detr, yolos

github-actions · 2026-08-18T17:23:57Z

Workflow Run ⚙️

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

models: ["models/conditional_detr", "models/deformable_detr", "models/detr", "models/grounding_dino", "models/rf_detr", "models/rt_detr", "models/yolos"]
quantizations: []

github-actions · 2026-08-18T17:49:57Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	9806cd79	workflow commit (merge commit)
PR	1f359f14	branch commit (from PR)
main	f9b76f2d	base commit (on `main`)

Model CI Report

❌ 2 new failed tests from this PR 😭

rf_detr:
tests/models/rf_detr/test_modeling_rf_detr.py::RfDetrModelIntegrationTest::test_inference_object_detection (✅ ⟹ ❌)
tests/models/rf_detr/test_modeling_rf_detr.py::RfDetrModelIntegrationTest::test_inference_segmentation (✅ ⟹ ❌)

guarin · 2026-08-19T09:29:53Z

run-slow: rf_detr

guarin · 2026-08-19T09:30:34Z

-                ("cuda", (8, 0)): [0.959521, 0.931229, 0.897797, 0.7286, 0.672863],
-                ("xpu", None): [0.959521, 0.931229, 0.897797, 0.7286, 0.672863],
+                ("cuda", (8, 0)): [0.960207, 0.934029, 0.896868, 0.718928, 0.707891],
+                ("cpu", None): [0.960285, 0.934246, 0.896883, 0.718877, 0.707271],


regenerated because of antialiasing change in preprocessing

github-actions · 2026-08-19T09:31:08Z

Workflow Run ⚙️

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

models: ["models/rf_detr"]
quantizations: []

github-actions · 2026-08-19T09:44:16Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	de878630	workflow commit (merge commit)
PR	8c1bf95b	branch commit (from PR)
main	e7e8b7fa	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

zucchini-nlp

It's great to see grouped iter (could be an mlinter rule 🤔 ). I am seeing a breaking changes, and not sure why are they added, left comments 👇🏻

zucchini-nlp · 2026-08-24T08:15:39Z

+            annotations = reorder_images(grouped_annotations, grouped_images_index)
            encoded_inputs["labels"] = [
                BatchFeature(annotation, tensor_type=return_tensors) for annotation in annotations
            ]


what is the diff here from DetrImageProcessor._preprocess?

rfdetr does rescale then resize while detr does resize then rescale

zucchini-nlp · 2026-08-24T08:26:06Z

+            if paired_list is None
+            else {key: [item] for key, item in _iterate_items(paired_list, is_nested)}


why we had to return a list[item] instead of unsqueezing a new dim?

because with disable_grouping==False we always returned a list instead of unsqueeze. The reason is that paired values are not necessarily tensors.

zucchini-nlp · 2026-08-24T08:34:04Z

-                    )
-                image = resized_image
+                orig_size = stacked_images.shape[-2:]
+                stacked_images = self.resize(stacked_images, size=size, resample=resample, antialias=False)


this is breaking, why we need to change antialias=False?

This if a fix from rf-detr: https://github.com/roboflow/rf-detr/blob/5547a9360c19df7ff8cedcd29f841f7de01f2f75/src/rfdetr/detr.py#L2480-L2481

oke, can we mark PR as 🚨

zucchini-nlp

LGTM, lets run slow tests for rf_detr before merging

zucchini-nlp · 2026-08-27T08:41:52Z

+            grouped_annotations[key] = stacked_annotations
+        processed_images = reorder_images(grouped_images, grouped_images_index)

        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)):


do_pad can go under the same loop since we're reusing the rescaled images without "reorder then group again" code in between

Sadly not possible because of padded_size = get_max_height_width(images) which requires the first loop to be complete.

guarin · 2026-09-15T14:39:53Z

run-slow: conditional_detr, deformable_detr, detr, grounding_dino, rf_detr, rt_detr, yolos

github-actions · 2026-09-15T14:40:58Z

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

run-slow: conditional_detr, deformable_detr, detr, grounding_dino, rf_detr, rt_detr, yolos

github-actions · 2026-09-15T14:58:24Z

CI recap

Dashboard: View test results in Grafana
Latest run: 34983156306:1
Result: success | Jobs: 16 | Tests: 96,401 | Failures: 0 | Duration: 9h 46m

molbap

Went through it! sounds good to me as well. Let's merge soon? maybe a quick core eye from @vasqu ?

vasqu

Imo just one design question but if you feel like it's not really appropriate feel free to ignore

guarin · 2026-09-21T16:38:45Z

run-slow: conditional_detr

github-actions · 2026-09-21T16:40:06Z

Nvidia CI

Workflow Run ⚙️

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

models: ["models/conditional_detr"]
quantizations: []

github-actions · 2026-09-21T16:40:08Z

AMD CI

Workflow Run ⚙️

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

models: ["models/conditional_detr"]

github-actions · 2026-09-21T16:45:34Z

CI Results (AMD)

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	f3b86ec1	workflow commit (merge commit)
PR	50c18824	branch commit (from PR)
main	d67c7293	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

github-actions · 2026-09-21T16:49:31Z

CI Results (Nvidia)

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	f3b86ec1	workflow commit (merge commit)
PR	50c18824	branch commit (from PR)
main	d67c7293	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

guarin added 4 commits August 18, 2026 16:48

speed up detr image processing

8367289

Merge branch 'main' into speed-up-detr-image-preprocessing

430fc25

clarify code

000593a

Merge branch 'speed-up-detr-image-preprocessing' of https://github.co…

e174bb3

…m/guarin/transformers into speed-up-detr-image-preprocessing

fix

1f359f1

update expectations

8c1bf95

guarin commented Aug 19, 2026

View reviewed changes

Comment thread tests/models/rf_detr/test_modeling_rf_detr.py

guarin commented Aug 19, 2026

View reviewed changes

This was referenced Aug 21, 2026

Simplify group_images_by_shape #48188

Closed

Simplify group images by shape #48192

Draft

zucchini-nlp reviewed Aug 24, 2026

View reviewed changes

guarin added 2 commits August 24, 2026 11:57

revert

fb1f314

add xpu values back

1e2496f

guarin changed the title ~~Speed up detr image processing~~ 🚨 Speed up detr image processing Aug 24, 2026

zucchini-nlp approved these changes Aug 27, 2026

View reviewed changes

Merge branch 'main' into speed-up-detr-image-preprocessing

584f498

update comment

50c1882

molbap approved these changes Sep 21, 2026

View reviewed changes

vasqu approved these changes Sep 21, 2026

View reviewed changes

Comment thread src/transformers/models/detr/image_processing_detr.py

Comment thread src/transformers/models/detr/image_processing_detr.py

Comment thread tests/models/rf_detr/test_modeling_rf_detr.py

		if paired_list is None
		else {key: [item] for key, item in _iterate_items(paired_list, is_nested)}

Conversation

guarin commented Aug 18, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Uh oh!

HuggingFaceDocBuilderDev commented Aug 18, 2026

Uh oh!

guarin commented Aug 18, 2026

Uh oh!

github-actions Bot commented Aug 18, 2026

Uh oh!

github-actions Bot commented Aug 18, 2026

CI Results

Commit Info

Model CI Report

Uh oh!

guarin commented Aug 19, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Aug 19, 2026

Uh oh!

github-actions Bot commented Aug 19, 2026

CI Results

Commit Info

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!

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

guarin commented Sep 15, 2026

Uh oh!

github-actions Bot commented Sep 15, 2026

Uh oh!

github-actions Bot commented Sep 15, 2026

CI recap

Uh oh!

molbap left a comment

Choose a reason for hiding this comment

Uh oh!

vasqu left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

guarin commented Sep 21, 2026

Uh oh!

github-actions Bot commented Sep 21, 2026

Nvidia CI

Uh oh!

github-actions Bot commented Sep 21, 2026

AMD CI

Uh oh!

github-actions Bot commented Sep 21, 2026

CI Results (AMD)

guarin commented Aug 18, 2026 •

edited by github-actions Bot

Loading