Flux.1 image generation has a relatively unfinished image output that inferenced on iGPU than CPU

[JamieVC]

Describe the bug
Flux.1 image generation has a relatively unfinished image output that inferenced on iGPU than CPU.

Config for iGPU inference and the output:

from optimum.intel.openvino import OVDiffusionPipeline
model_dir = model_base_dir / "INT4" if use_quantized_models.value else model_base_dir / "FP16"
ov_config = {"INFERENCE_PRECISION_HINT" : "f16", "ACTIVATIONS_SCALE_FACTOR": "8.0", "CACHE_DIR" : ""}
ov_pipe = OVDiffusionPipeline.from_pretrained(model_dir, device=device.value, ov_config=ov_config)

Config for CPU inference and the output:

from optimum.intel.openvino import OVDiffusionPipeline
model_dir = model_base_dir / "INT4" if use_quantized_models.value else model_base_dir / "FP16" 
ov_pipe = OVDiffusionPipeline.from_pretrained(model_dir, device=device.value)

Expected behavior
Output that inferenced on iGPU should also complete a sign with the whole texts.

Installation instructions (Please mark the checkbox)
[ yes ] I followed the installation guide at https://github.com/openvinotoolkit/openvino_notebooks#-installation-guide to install the notebooks.

[dnkurek]

What version of OpenVINO are you using?

[geunhwan]

This is known issue to be fixed. We seem to have an issue in sdpa. (cc. @e-ddykim)

[JamieVC]

onnx 1.17.0
openai 1.57.1
opencv-python 4.10.0.84
openvino 2025.0.0 (local Build with the PR #27265 )
openvino-genai 2024.6.0.0rc3
openvino-telemetry 2024.1.0
openvino-tokenizers 2024.6.0.0
optimum 1.24.0.dev0
optimum-intel 1.22.0.dev0+58aec63

[JamieVC]

sdpa

hi @geunhwan
Can I know what issue# in internal JIRA system?

[e-ddykim]

Hi @JamieVC, #28828 is ready for review.