Latent Consistency Model using Optimum-Intel OpenVINO#

This Jupyter notebook can be launched after a local installation only.

Github

This notebook provides instructions how to run Latent Consistency Model (LCM). It allows to setup standard Hugging Face diffusers pipeline and Optimum Intel pipeline optimized for Intel hardware including CPU and GPU. Running inference on CPU and GPU it is easy to compare performance and time required to generate an image for provided prompt. The notebook can be also used on other Intel hardware with minimal or no modifications.

image0

Optimum Intel is an interface from Hugging Face between both diffusers and transformers libraries and various tools provided by Intel to accelerate pipelines on Intel hardware. It allows to perform quantization of the models hosted on Hugging Face. In this notebook OpenVINO is used for AI-inference acceleration as a backend for Optimum Intel!

For more details please refer to Optimum Intel repository huggingface/optimum-intel

LCMs are the next generation of generative models after Latent Diffusion Models (LDMs). They are proposed to overcome the slow iterative sampling process of Latent Diffusion Models (LDMs), enabling fast inference with minimal steps (from 2 to 4) on any pre-trained LDMs (e.g. Stable Diffusion). To read more about LCM please refer to https://latent-consistency-models.github.io/

Table of contents:#

Prerequisites#

Install required packages

%pip install -q "openvino>=2023.3.0"
%pip install -q "onnx>=1.11.0"
%pip install -q "optimum-intel[diffusers]@git+https://github.com/huggingface/optimum-intel.git" "ipywidgets" "torch>=2.1" "transformers>=4.33.0" --extra-index-url https://download.pytorch.org/whl/cpu
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.
import warnings

warnings.filterwarnings("ignore")

Showing Info Available Devices#

The available_devices property shows the available devices in your system. The “FULL_DEVICE_NAME” option to ie.get_property() shows the name of the device. Check what is the ID name for the discrete GPU, if you have integrated GPU (iGPU) and discrete GPU (dGPU), it will show device_name="GPU.0" for iGPU and device_name="GPU.1" for dGPU. If you just have either an iGPU or dGPU that will be assigned to "GPU"

Note: For more details about GPU with OpenVINO visit this link. If you have been facing any issue in Ubuntu 20.04 or Windows 11 read this blog.

import openvino as ov

core = ov.Core()
devices = core.available_devices

for device in devices:
    device_name = core.get_property(device, "FULL_DEVICE_NAME")
    print(f"{device}: {device_name}")
CPU: Intel(R) Core(TM) i9-10920X CPU @ 3.50GHz

Using full precision model in CPU with LatentConsistencyModelPipeline#

Standard pipeline for the Latent Consistency Model(LCM) from Diffusers library is used here. For more information please refer to https://huggingface.co/docs/diffusers/en/api/pipelines/latent_consistency_models

from diffusers import LatentConsistencyModelPipeline
import gc

pipeline = LatentConsistencyModelPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7")
2024-06-06 00:42:53.373765: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2024-06-06 00:42:53.408053: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-06-06 00:42:53.902704: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
prompt = "A cute squirrel in the forest, portrait, 8k"

image = pipeline(prompt=prompt, num_inference_steps=4, guidance_scale=8.0, height=512, width=512).images[0]
image.save("image_standard_pipeline.png")
image
0%|          | 0/4 [00:00<?, ?it/s]
../_images/latent-consistency-models-optimum-demo-with-output_8_1.png
del pipeline
gc.collect();

Select inference device for text-to-image generation#

import ipywidgets as widgets

core = ov.Core()

device = widgets.Dropdown(
    options=core.available_devices + ["AUTO"],
    value="CPU",
    description="Device:",
    disabled=False,
)

device
Dropdown(description='Device:', options=('CPU', 'AUTO'), value='CPU')

Running inference using Optimum Intel OVLatentConsistencyModelPipeline#

Accelerating inference of LCM using Intel Optimum with OpenVINO backend. For more information please refer to https://huggingface.co/docs/optimum/intel/inference#latent-consistency-models. The pretrained model in this notebook is available on Hugging Face in FP32 precision and in case if CPU is selected as a device, then inference runs with full precision. For GPU accelerated AI-inference is supported for FP16 data type and FP32 precision for GPU may produce high memory footprint and latency. Therefore, default precision for GPU in OpenVINO is FP16. OpenVINO GPU Plugin converts FP32 to FP16 on the fly and there is no need to do it manually

from optimum.intel.openvino import OVLatentConsistencyModelPipeline
from pathlib import Path

if not Path("./openvino_ir").exists():
    ov_pipeline = OVLatentConsistencyModelPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", height=512, width=512, export=True, compile=False)
    ov_pipeline.save_pretrained("./openvino_ir")
else:
    ov_pipeline = OVLatentConsistencyModelPipeline.from_pretrained("./openvino_ir", export=False, compile=False)

ov_pipeline.reshape(batch_size=1, height=512, width=512, num_images_per_prompt=1)
INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino
Framework not specified. Using pt to export the model.
Keyword arguments {'subfolder': '', 'token': None, 'trust_remote_code': False} are not expected by StableDiffusionPipeline and will be ignored.
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
Using framework PyTorch: 2.3.1+cpu
WARNING:tensorflow:Please fix your imports. Module tensorflow.python.training.tracking.base has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.
[ WARNING ]  Please fix your imports. Module %s has been moved to %s. The old module will be deleted in version %s.
Using framework PyTorch: 2.3.1+cpu
Using framework PyTorch: 2.3.1+cpu
Using framework PyTorch: 2.3.1+cpu
OVLatentConsistencyModelPipeline {
  "_class_name": "OVLatentConsistencyModelPipeline",
  "_diffusers_version": "0.24.0",
  "feature_extractor": [
    "transformers",
    "CLIPImageProcessor"
  ],
  "requires_safety_checker": true,
  "safety_checker": [
    "stable_diffusion",
    "StableDiffusionSafetyChecker"
  ],
  "scheduler": [
    "diffusers",
    "LCMScheduler"
  ],
  "text_encoder": [
    "optimum",
    "OVModelTextEncoder"
  ],
  "text_encoder_2": [
    null,
    null
  ],
  "tokenizer": [
    "transformers",
    "CLIPTokenizer"
  ],
  "unet": [
    "optimum",
    "OVModelUnet"
  ],
  "vae_decoder": [
    "optimum",
    "OVModelVaeDecoder"
  ],
  "vae_encoder": [
    "optimum",
    "OVModelVaeEncoder"
  ]
}
ov_pipeline.to(device.value)
ov_pipeline.compile()
Compiling the vae_decoder to CPU ...
Compiling the unet to CPU ...
Compiling the vae_encoder to CPU ...
Compiling the text_encoder to CPU ...
prompt = "A cute squirrel in the forest, portrait, 8k"

image_ov = ov_pipeline(prompt=prompt, num_inference_steps=4, guidance_scale=8.0, height=512, width=512).images[0]
image_ov.save("image_opt.png")
image_ov
0%|          | 0/4 [00:00<?, ?it/s]
../_images/latent-consistency-models-optimum-demo-with-output_15_1.png
del ov_pipeline
gc.collect();