Distributed Services#

BentoML provides a flexible framework for deploying machine learning models as Services. While a single Service often suffices for most use cases, it is useful to create multiple Services running in a distributed way in more complex scenarios.

This document provides guidance on creating and deploying a BentoML project with distributed Services.

Single and distributed Services#

Using a single BentoML Service in service.py is typically sufficient for most use cases. This approach is straightforward, easy to manage, and works well when you only need to deploy a single model and the API logic is simple.

In deployment, a BentoML Service run as processes in a container. If you define multiple Services, they run as processes in different containers. This distributed approach is useful when dealing with more complex scenarios, such as:

Pipelining CPU and GPU processing for better throughput: Distributing tasks between CPU and GPU can enhance throughput. Certain preprocessing or postprocessing tasks might be more efficiently handled by the CPU, while the GPU focuses on model inference.
Optimizing resource utilization and scalability: Distributed Services can run on different instances, allowing for independent scaling and efficient resource usage. This flexibility is important in handling varying loads and optimizing specific resource demands.
Asymmetrical GPU requirements: Different models might have varied GPU requirements. Distributing these models across Services helps you allocate resources more efficiently and cost-effectively.
Handling complex workflows: For applications involving intricate workflows, like sequential processing, parallel processing, or the composition of multiple models, you can create multiple Services to modularize these processes if necessary, improving maintainability and efficiency.

Interservice communication#

Distributed Services support complex, modular architectures through interservice communication. Different Services can interact with each other using the bentoml.depends() function. This allows for direct method calls between Services as if they were local class functions. Key features of interservice communication:

Automatic service discovery & routing: When Services are deployed, BentoML handles the discovery of Services, routes requests appropriately, and manages payload serialization and deserialization.
Arbitrary dependency chains: Services can form dependency chains of any length, enabling intricate Service orchestration.
Diamond-shaped dependencies: Support a diamond dependency pattern, where multiple Services depend on a single downstream Service, for maximizing Service reuse.

The following is an example of two distributed Services with different hardware requirements and one Service depends on another using bentoml.depends().

SDXLControlNetService: A high-resource demanding Service, requiring GPU support for image generation.
ControlNet: A Service designed to handle incoming requests and route them appropriately. It calls a method of the SDXLControlNetService Service.

This is the service.py file:

from __future__ import annotations

import typing as t

import cv2
import numpy as np
import PIL
from PIL.Image import Image as PIL_Image

import torch
from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, AutoencoderKL
from pydantic import BaseModel

import bentoml

CONTROLNET_MODEL_ID = "diffusers/controlnet-canny-sdxl-1.0"
VAE_MODEL_ID = "madebyollin/sdxl-vae-fp16-fix"
BASE_MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"


@bentoml.service(
    traffic={"timeout": 600},
    workers=1,
    resources={
        "gpu": "1",
        "gpu_type": "nvidia-l4",
    }
)
class SDXLControlNetService:

    def __init__(self) -> None:

        if torch.cuda.is_available():
            self.device = "cuda"
            self.dtype = torch.float16
        else:
            self.device = "cpu"
            self.dtype = torch.float32

        self.controlnet = ControlNetModel.from_pretrained(
            CONTROLNET_MODEL_ID,
            torch_dtype=self.dtype,
        )

        self.vae = AutoencoderKL.from_pretrained(
            VAE_MODEL_ID,
            torch_dtype=self.dtype,
        )

        self.pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
            BASE_MODEL_ID,
            controlnet=self.controlnet,
            vae=self.vae,
            torch_dtype=self.dtype
        ).to(self.device)


    @bentoml.api
    async def generate(
            self,
            prompt: str,
            arr: np.ndarray[t.Any, np.uint8],
            **kwargs,
    ):
        image = PIL.Image.fromarray(arr)
        return self.pipe(prompt, image=image, **kwargs).to_tuple()


class Params(BaseModel):
    prompt: str
    negative_prompt: t.Optional[str]
    controlnet_conditioning_scale: float = 0.5
    num_inference_steps: int = 25


@bentoml.service(
    traffic={"timeout": 600},
    workers=8,
resources={"cpu": "1"}
)
class ControlNet:
    # Pass the dependent Service class as an argument
    controlnet_service = bentoml.depends(SDXLControlNetService)

    @bentoml.api
    async def generate(self, image: PIL_Image, params: Params) -> PIL_Image:
        arr = np.array(image)
        arr = cv2.Canny(arr, 100, 200)
        arr = arr[:, :, None]
        arr = np.concatenate([arr, arr, arr], axis=2)
        params_d = params.dict()
        prompt = params_d.pop("prompt")
        # Invok a class level function of another Service
        res = await self.controlnet_service.generate(
            prompt,
            arr=arr,
            **params_d
        )
        return res[0][0]

To declare a dependency, you use the bentoml.depends() function by passing the dependent Service class as an argument. This creates a direct link between Services, facilitating easy method invocation. This example uses the following code to achieve this:

class ControlNet:
    controlnet_service = bentoml.depends(SDXLControlNetService)

Once a dependency is declared, invoking methods on the dependent Service is similar to calling a local method. In other words, Service A can call Service B as if Service A were invoking a class level function on Service B. This abstracts away the complexities of network communication, serialization, and deserialization. In this example, the Service ControlNet invokes the generate function of SDXLControlNetService as below:

res = await self.controlnet_service.generate(prompt, arr=arr, **params_d)

Using bentoml.depends() is a recommended way for creating a BentoML project with distributed Services. It enhances modularity as you can develop reusable, loosely coupled Services that can be maintained and scaled independently.

`bentofile.yaml`#

For projects with multiple Services, you should reference the primary Service handling user requests for the service field in bentofile.yaml. For example:

service: "service:ControlNet" #ControlNet is the one that receives users' requests
labels:
  owner: bentoml-team
  project: gallery
...

You can then containerize it as an OCI-compliant image or deploy it to BentoCloud.

Deploy distributed Services#

Deploying a project with distributed Services to BentoCloud is similar to deploying a single Service, with nuances in setting custom configurations.

To set custom configurations for each, we recommend you use a separate configuration file and reference it in the BentoML CLI command or Python API for deployment.

The following is an example file that defines some custom configurations for the above two Services. You set configurations of each Service in the services field. Refer to Configure Deployments to see the available configuration fields.

# config-file.yaml
name: "deployment-name"
description: "This project creates an image generation application based on users' requirements."
envs: # Optional. If you specify environment variables here, they will be applied to all Services
  - name: "GLOBAL_ENV_VAR_NAME"
    value: "env_var_value"
services: # Add the configs of each Service under this field
  SDXLControlNetService: # Service one
    instance_type: "gpu.l4.1"
    scaling:
      max_replicas: 2
      min_replicas: 1
    envs: # Environment variables specific to Service one
      - name: "ENV_VAR_NAME"
        value: "env_var_value"
    deployment_strategy: "RollingUpdate"
    config_overrides:
      traffic:
        # float in seconds
        timeout: 700
        max_concurrency: 20
        external_queue: true
      resources:
        cpu: "400m"
        memory: "1Gi"
      workers:
        - gpu: 1
  ControlNet: # Service two
    instance_type: "cpu.1"
    scaling:
      max_replicas: 5
      min_replicas: 1

To deploy this Service to BentoCloud, you can choose either the BentoML CLI or Python API:

BentoML CLI

bentoml deploy . -f config-file.yaml

Python API

import bentoml
bentoml.deployment.create(bento = "./path_to_your_project", config_file="config-file.yaml")

Distributed Services#

Single and distributed Services#

Interservice communication#

bentofile.yaml#

Deploy distributed Services#

`bentofile.yaml`#