​
Our Observations

We have deployed this model using A100 GPU and observed that the model took an average cold start time of 7.02 sec and an average inference time of 34 sec for generating a video of 4 sec with 6fps

​
Defining Dependencies

We are using the HuggingFace Diffusers library for the deployment. You can also use this script for deployment provided by Stability AI.

​
Constructing the GitHub/GitLab Template

Now quickly construct the GitHub/GitLab template, this process is mandatory and make sure you don’t add any file named model.py

stable-video-diffusion/
├── app.py
└── config.yaml

You can also add other files to this directory.

​
Create the class for inference

In the app.py we will define the class and import all the required functions

  1. def initialize: In this function, you will initialize your model and define any variable that you want to use during inference. We are using torch.compile which improves the latency but requires a large GPU(A10/A100). If you are using Nvidia T4 then remove those lines and use model CPU offloading to reduce memory usage.
self.pipe.enable_model_cpu_offload()

class InferlessPythonModel:
    
    def initialize(self):
        self.pipe = StableVideoDiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16")
        self.pipe.to("cuda")
        self.pipe.unet = torch.compile(self.pipe.unet, mode="reduce-overhead", fullgraph=True)
        self.pipe.vae = torch.compile(self.pipe.vae, mode="reduce-overhead", fullgraph=True)
        self.seed = 42
        self.randomize_seed = True
        self.motion_bucket_id = 127
        self.fps_id = 6
        self.version = "svd_xt"
        self.cond_aug = 0.02,
        self.decoding_t = 3
        self.device = "cuda"
        self.output_folder = "outputs"
  1. def resize: This is a utility function required during inference, it will download the image and then resize it. You can also add any other functions which are required during inference.
    def resize_image(self,image_url, output_size=(1024, 576)):
        # Calculate aspect ratio
        response = requests.get(image_url)
        image_data = BytesIO(response.content)

        # Open the image using PIL
        image = Image.open(image_data)
        target_aspect = output_size[0] / output_size[1]  # Aspect ratio of the desired size
        image_aspect = image.width / image.height  # Aspect ratio of the original image

        # Resize then crop if the original image is larger
        if image_aspect > target_aspect:
            # Resize the image to match the target height, maintaining the aspect ratio
            new_height = output_size[1]
            new_width = int(new_height * image_aspect)
            resized_image = image.resize((new_width, new_height), Image.LANCZOS)
            # Calculate coordinates for cropping
            left = (new_width - output_size[0]) / 2
            top = 0
            right = (new_width + output_size[0]) / 2
            bottom = output_size[1]
        else:
            # Resize the image to match the target width, maintaining the aspect ratio
            new_width = output_size[0]
            new_height = int(new_width / image_aspect)
            resized_image = image.resize((new_width, new_height), Image.LANCZOS)
            # Calculate coordinates for cropping
            left = 0
            top = (new_height - output_size[1]) / 2
            right = output_size[0]
            bottom = (new_height + output_size[1]) / 2

        # Crop the image
        cropped_image = resized_image.crop((left, top, right, bottom))
        return cropped_image
  1. def infer: This function gets called for every request that you send. Here you can define all the steps that are required for the inference. You can also pass custom values for inference, for example fps_id is fixed in the tutorial, you can pass it through inputs parameter.
    def infer(self,inputs):
        image_url = inputs['image_url']
        max_64_bit_int = 2**63 - 1
        image = self.resize_image(image_url)

        if image.mode == "RGBA":
            image = image.convert("RGB")

        if(self.randomize_seed):
            seed = random.randint(0, max_64_bit_int)
        generator = torch.manual_seed(seed)

        os.makedirs(self.output_folder, exist_ok=True)
        base_count = len(glob(os.path.join(self.output_folder, "*.mp4")))
        video_path = os.path.join(self.output_folder, f"{base_count:06d}.mp4")

        frames = self.pipe(image, decode_chunk_size=self.decoding_t, generator=generator, motion_bucket_id=self.motion_bucket_id, noise_aug_strength=0.1).frames[0]
        export_to_video(frames, video_path, fps=self.fps_id)
        torch.manual_seed(seed)

        #Video convert to base64
        with open(video_path, "rb") as video_file:
            video_binary_data = video_file.read()
            video_bytes_io = BytesIO(video_binary_data)
            base64_encoded_data = base64.b64encode(video_bytes_io.read())
            base64_string = base64_encoded_data.decode("utf-8")

        return {"generated_video": base64_string}
  1. def finalize: This function cleans up all the allocated memory.
    def finalize(self,*args):
        self.pipe = None

​
Creating the Custom Runtime

This is a mandatory step where we allow the users to upload their own custom runtime through config.yaml.

build:
  cuda_version: "12.1.1"
  system_packages:
    - "libssl-dev"
    - "libx11-6"
    - "libxext6"
    - "libgl1-mesa-glx"
  python_packages:
    - "accelerate==0.25.0"
    - "opencv-python==4.8.1.78"
    - "git+https://github.com/huggingface/diffusers.git@refs/pull/6103/head"
    - "transformers==4.35.2"
    - "requests==2.31.0"

​
Deploying the model on Inferless

Inferless supports multiple ways of importing your model. For this tutorial, we will use GitHub.

​
Import the Model through GitHub

Click on theRepo(Custom code) and then click on the Add provider to connect to your GitHub account. Once your account integration is completed, click on your Github account and continue.

​
Provide the Model details

Enter the name of the model and pass the GitHub repository URL.

Now, you can define the model’s input and output parameters in JSON format. For more information, go through this page.

  • INPUT: Refer to the function def infer(self, inputs), here we mentioned inputs["image_url"], which means inputs the parameter will have a key prompt.

  • OUTPUT: The same goes here, the function mentioned above will return the results as a key-value pair return {"generated_video": result}.

Input JSON must include prompt a key to pass the instruction. For output JSON, it must be includedgenerated_result to retrieve the output.

​
GPU Selection

On the Inferless platform, we support all the GPUs. For this tutorial, we recommend using A100 GPU. Select A100 from the drop-down menu in the GPU Type.

​
Using the Custom Runtime

In the Advance configuration, we have the option to select the custom runtime. First, click on the Add runtime to upload the config.yaml file, give any name and save it. Choose the runtime from the drop-down menu and then click on continue.

​
Review and Deploy

In this final stage, carefully review all modifications. Once you’ve examined all changes, proceed to deploy the model by clicking the Import button.

Voilà, your model is now deployed!

Note: This demo GIF shows the deployment of the Stable Video Diffusion model using webhook.

Deploy a CodeLlama-Python-34B Model using InferlessDeploy Mixtral-8x7B using Inferless