Skip to content

[OpenVINO]Add Kontext support #1374

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 17 commits into
base: main
Choose a base branch
from
Open

[OpenVINO]Add Kontext support #1374

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
17 commits
Select commit Hold shift + click to select a range
ebe5be2
add kontext support
openvino-dev-samples Jul 7, 2025
d2a5a92
add kontext support
openvino-dev-samples Jul 7, 2025
024ac2d
add test case
openvino-dev-samples Jul 7, 2025
c365139
reformat
openvino-dev-samples Jul 7, 2025
051a8da
Apply style fixes
github-actions[bot] Jul 7, 2025
c49252f
add diffusion and export test
openvino-dev-samples Jul 8, 2025
8459565
rebase the export cli test
openvino-dev-samples Jul 9, 2025
2c7e1de
add quantize test case
openvino-dev-samples Jul 9, 2025
c160e7f
delete cli export test
openvino-dev-samples Jul 9, 2025
0d5d3ab
Merge branch 'huggingface:main' into kontext
openvino-dev-samples Aug 20, 2025
1e37d3d
support diffusers 3.5
openvino-dev-samples Aug 20, 2025
cb7df16
update diffusion test with static shape
openvino-dev-samples Sep 10, 2025
8b8e2c8
remove quantization test
openvino-dev-samples Sep 10, 2025
dd534af
add flux-kontext to _HEAD_TO_AUTOMODELS
openvino-dev-samples Sep 18, 2025
bcd5070
update for kontext
openvino-dev-samples Sep 18, 2025
4ae8de2
update int8 test case
openvino-dev-samples Sep 27, 2025
7a20026
solve conflict
openvino-dev-samples Oct 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions optimum/intel/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,7 @@
"OVFluxPipeline",
"OVFluxImg2ImgPipeline",
"OVFluxInpaintPipeline",
"OVFluxKontextPipeline",
"OVFluxFillPipeline",
"OVSanaPipeline",
"OVPipelineForImage2Image",
Expand Down Expand Up @@ -162,6 +163,7 @@
"OVFluxImg2ImgPipeline",
"OVFluxInpaintPipeline",
"OVFluxFillPipeline",
"OVFluxKontextPipeline",
"OVSanaPipeline",
"OVPipelineForImage2Image",
"OVPipelineForText2Image",
Expand Down
1 change: 1 addition & 0 deletions optimum/intel/openvino/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,7 @@
OVFluxFillPipeline,
OVFluxImg2ImgPipeline,
OVFluxInpaintPipeline,
OVFluxKontextPipeline,
OVFluxPipeline,
OVLatentConsistencyModelImg2ImgPipeline,
OVLatentConsistencyModelPipeline,
Expand Down
16 changes: 16 additions & 0 deletions optimum/intel/openvino/modeling_diffusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,6 +121,11 @@
else:
SanaSprintPipeline = object

if is_diffusers_version(">=", "0.35.0"):
from diffusers import FluxKontextPipeline
else:
FluxKontextPipeline = object


if is_diffusers_version(">=", "0.35.0"):
from diffusers.models.cache_utils import CacheMixin
Expand Down Expand Up @@ -1676,6 +1681,12 @@ class OVFluxFillPipeline(OVDiffusionPipeline, OVTextualInversionLoaderMixin, Flu
auto_model_class = FluxFillPipeline


class OVFluxKontextPipeline(OVDiffusionPipeline, OVTextualInversionLoaderMixin, FluxKontextPipeline):
main_input_name = "image"
export_feature = "image-to-image"
auto_model_class = FluxKontextPipeline


class OVSanaPipeline(OVDiffusionPipeline, OVTextualInversionLoaderMixin, SanaPipeline):
main_input_name = "prompt"
export_feature = "text-to-image"
Expand Down Expand Up @@ -1779,6 +1790,11 @@ def _get_ov_class(pipeline_class_name: str, throw_error_if_not_exist: bool = Tru
SUPPORTED_OV_PIPELINES.append(OVSanaSprintPipeline)
OV_TEXT2IMAGE_PIPELINES_MAPPING["sana-sprint"] = OVSanaSprintPipeline


if is_diffusers_version(">=", "0.34.0"):
SUPPORTED_OV_PIPELINES.append(OVFluxKontextPipeline)
OV_IMAGE2IMAGE_PIPELINES_MAPPING["flux-kontext"] = OVFluxKontextPipeline

SUPPORTED_OV_PIPELINES_MAPPINGS = [
OV_TEXT2IMAGE_PIPELINES_MAPPING,
OV_IMAGE2IMAGE_PIPELINES_MAPPING,
Expand Down
1 change: 1 addition & 0 deletions optimum/intel/openvino/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,6 +139,7 @@
"sana": "OVSanaPipeline",
"flux": "OVFluxPipeline",
"flux-fill": "OVFluxFillPipeline",
"flux-kontext": "OVFluxKontextPipeline",
"pix2struct": "OVModelForPix2Struct",
"latent-consistency": "OVLatentConsistencyModelPipeline",
"open_clip_text": "OVModelOpenCLIPText",
Expand Down
11 changes: 11 additions & 0 deletions optimum/intel/utils/dummy_openvino_and_diffusers_objects.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,6 +246,17 @@ def from_pretrained(cls, *args, **kwargs):
requires_backends(cls, ["openvino", "diffusers"])


class OVFluxKontextPipeline(metaclass=DummyObject):
_backends = ["openvino", "diffusers"]

def __init__(self, *args, **kwargs):
requires_backends(self, ["openvino", "diffusers"])

@classmethod
def from_pretrained(cls, *args, **kwargs):
requires_backends(cls, ["openvino", "diffusers"])


class OVSanaPipeline(metaclass=DummyObject):
_backends = ["openvino", "diffusers"]

Expand Down
36 changes: 26 additions & 10 deletions tests/openvino/test_diffusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -497,6 +497,8 @@ class OVPipelineForImage2ImageTest(unittest.TestCase):
"stable-diffusion-3",
"flux",
]
if is_diffusers_version(">=", "0.35.0"):
SUPPORTED_ARCHITECTURES.append("flux-kontext")
AUTOMODEL_CLASS = AutoPipelineForImage2Image
OVMODEL_CLASS = OVPipelineForImage2Image
TASK = "image-to-image"
Expand All @@ -508,11 +510,12 @@ def generate_inputs(self, height=128, width=128, batch_size=1, channel=3, input_
height=height, width=width, batch_size=batch_size, channel=channel, input_type=input_type
)

if model_type in ["flux", "stable-diffusion-3"]:
if model_type in ["flux", "stable-diffusion-3", "flux-kontext"]:
inputs["height"] = height
inputs["width"] = width

inputs["strength"] = 0.75

if model_type != "flux-kontext":
inputs["strength"] = 0.75

return inputs

Expand Down Expand Up @@ -544,7 +547,16 @@ def test_num_images_per_prompt(self, model_arch: str):
height=height, width=width, batch_size=batch_size, model_type=model_arch
)
outputs = pipeline(**inputs, num_images_per_prompt=num_images_per_prompt).images
self.assertEqual(outputs.shape, (batch_size * num_images_per_prompt, height, width, 3))
if model_arch != "flux-kontext":
self.assertEqual(outputs.shape, (batch_size * num_images_per_prompt, height, width, 3))
else:
# output shape is fixed: https://github.com/huggingface/diffusers/blob/v0.35.1/src/diffusers/pipelines/flux/pipeline_flux_kontext.py#L882
if (height == width):
self.assertEqual(outputs.shape, (batch_size * num_images_per_prompt, 1024, 1024, 3))
elif (height > width):
self.assertEqual(outputs.shape, (batch_size * num_images_per_prompt, 1448, 724, 3))
else:
self.assertEqual(outputs.shape, (batch_size * num_images_per_prompt, 724, 1448, 3))

@parameterized.expand(["stable-diffusion", "stable-diffusion-xl", "latent-consistency"])
@require_diffusers
Expand Down Expand Up @@ -577,8 +589,11 @@ def __call__(self, *args, **kwargs) -> None:
@require_diffusers
def test_shape(self, model_arch: str):
pipeline = self.OVMODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch])

height, width, batch_size = 128, 64, 1
if model_arch != "flux-kontext":
# output shape is fixed: https://github.com/huggingface/diffusers/blob/v0.35.1/src/diffusers/pipelines/flux/pipeline_flux_kontext.py#L882
height, width, batch_size = 128, 64, 1
else:
height, width, batch_size = 1448, 724, 1

for input_type in ["pil", "np", "pt"]:
inputs = self.generate_inputs(
Expand All @@ -595,7 +610,7 @@ def test_shape(self, model_arch: str):
elif output_type == "pt":
self.assertEqual(outputs.shape, (batch_size, 3, height, width))
else:
if model_arch != "flux":
if not model_arch.startswith("flux"):
out_channels = (
pipeline.unet.config.out_channels
if pipeline.unet is not None
Expand All @@ -620,9 +635,10 @@ def test_shape(self, model_arch: str):
@require_diffusers
def test_compare_to_diffusers_pipeline(self, model_arch: str):
height, width, batch_size = 128, 128, 1
inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size, model_type=model_arch)

diffusers_pipeline = self.AUTOMODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch])
inputs = self.generate_inputs(height=height, width=width, batch_size=batch_size, model_type=model_arch)
auto_cls = self.AUTOMODEL_CLASS

diffusers_pipeline = auto_cls.from_pretrained(MODEL_NAMES[model_arch])
ov_pipeline = self.OVMODEL_CLASS.from_pretrained(MODEL_NAMES[model_arch])

for output_type in ["latent", "np", "pt"]:
Expand Down
4 changes: 4 additions & 0 deletions tests/openvino/test_exporters_cli.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@
from optimum.intel import ( # noqa
OVFluxFillPipeline,
OVFluxPipeline,
OVFluxKontextPipeline,
OVLatentConsistencyModelPipeline,
OVLTXPipeline,
OVModelForAudioClassification,
Expand Down Expand Up @@ -68,6 +69,7 @@
is_openvino_version,
is_tokenizers_version,
is_transformers_version,
is_diffusers_version,
)


Expand Down Expand Up @@ -98,6 +100,7 @@ class OVCLIExportTestCase(unittest.TestCase):
("text-to-image", "stable-diffusion-3"),
("text-to-image", "flux"),
("inpainting", "flux-fill"),
("image-to-image", "flux-kontext"),
("text-to-image", "sana"),
("text-to-video", "ltx-video"),
("feature-extraction", "sam"),
Expand All @@ -120,6 +123,7 @@ class OVCLIExportTestCase(unittest.TestCase):
"stable-diffusion-3": 6 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 2,
"flux": 4 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 0,
"flux-fill": 4 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 0,
"flux-kontext": 4 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 0,
"llava": 2 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 0,
"sana": 2 if is_tokenizers_version("<", "0.20.0") or is_openvino_version(">=", "2024.5") else 0,
"ltx-video": 2 if is_tokenizers_version("<", "0.20.0") or is_openvino_version(">=", "2024.5") else 0,
Expand Down
8 changes: 8 additions & 0 deletions tests/openvino/utils_tests.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,7 @@
"flaubert": "hf-internal-testing/tiny-random-flaubert",
"flux": "katuni4ka/tiny-random-flux",
"flux-fill": "katuni4ka/tiny-random-flux-fill",
"flux-kontext": "snake7gun/flux-kontext-random",
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe adding this here is not enough for the test for this model to be run. Could you please add changes to test_diffusion.py, test_exporters_cli.py and test_quantization.py similar to how it was done in this PR for example #1106?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @IlyasMoutawwakil my patch depends on the fix for diffusers lib, could you help to check ?
huggingface/diffusers#11875

Copy link
Member

@IlyasMoutawwakil IlyasMoutawwakil Jul 8, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

you don't need to wait for diffusers to do an entire release to test this. You can do the same as sana by using DiffusionPipeline instead of AutoPipelineForxxx https://github.com/huggingface/optimum-intel/blob/main/tests/openvino/test_diffusion.py#L146 for now

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

you don't need to wait for diffusers to do an entire release to test this. You can do the same as sana by using DiffusionPipeline instead of AutoPipelineForxxx https://github.com/huggingface/optimum-intel/blob/main/tests/openvino/test_diffusion.py#L146 for now

yes, but how about in CLI export test case ?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

cli export test should pass as it uses the correct class name https://github.com/huggingface/optimum/blob/ec756fdd00543d3a93458fafbcdf5abbb17b61ba/optimum/exporters/tasks.py#L2159 🙂

Copy link
Contributor Author

@openvino-dev-samples openvino-dev-samples Jul 10, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

cli export test should pass as it uses the correct class name https://github.com/huggingface/optimum/blob/ec756fdd00543d3a93458fafbcdf5abbb17b61ba/optimum/exporters/tasks.py#L2159 🙂

  1. CLI export should not work without my patch on diffusers
  2. another question is if there is a method to update the version of diffusers in pytest?
    E ImportError: cannot import name 'FluxKontextPipeline' from 'diffusers'

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. no it should work without it. your patch is only rleevant to the AutoPipelineForxxx class, not the DiffusionPipeline class.
  2. update it to what version ? I just checked the latest diffusers and this model is not even released in diffusers yet https://github.com/huggingface/diffusers/blob/v0.34.0/src/diffusers/pipelines/flux/pipeline_flux_kontext.py
    what's exactly the motive for adding this model anyway ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So your suggestion is adding the model after 0.34.0 version of diffuser released, am i right ?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

or install diffusers from source.. for testing sake.
but bear in mind that python packages can't be released with a "git+https...." dependency, so we will have to wait for a diffusers release if you want this in the next optimum-intel release.

"gpt_bigcode": "hf-internal-testing/tiny-random-GPTBigCodeModel",
"gpt2": "hf-internal-testing/tiny-random-gpt2",
"gpt_neo": "hf-internal-testing/tiny-random-GPTNeoModel",
Expand Down Expand Up @@ -269,6 +270,13 @@
"text_encoder": 64,
"text_encoder_2": 64,
},
"flux-kontext": {
"transformer": 60,
"vae_decoder": 30,
"vae_encoder": 26,
"text_encoder": 64,
"text_encoder_2": 76,
},
"llava": {
"lm_model": 30,
"text_embeddings_model": 1,
Expand Down