ENOVA

ENOVA is an open-source llm deployment, monitoring, injection and auto-scaling service. Since the increasing popularity of large language model (LLM) backend systems, it is common and necessary to deploy stable serverless serving of LLM on GPU clusters with auto-scaling. However, there exist challenges because the diversity and co-location of applications in GPU clusters will lead to low service quality and GPU utilization. To address them, we build ENOVA, which deconstructs the execution process of LLM service comprehensively and designs a configuration recommendation module for automatic deployment on any GPU clusters and a performance detection module for auto-scaling. On top of them, ENOVA implements a deployment execution engine for GPU cluster scheduling.

🤖 Architecture of ENOVA

✈️ Getting Started

Requirement

OS: Linux
Docker
Python: >=3.10
GPU: Nvidia GPUs with compute capability 7.0 or higher

Installation

Make sure you have Docker and Python Environment.
Use pip to install ENOVA:

# Create a new Python environment
conda create -n enova_env python=3.10
conda activate enova_env

# Install ENOVA
# Source: https://pypi.python.org/simple/
pip install enova_instrumentation_llmo
pip install enova

To verify the installation, run:

enova -h

Quickstart

Start an all-in-one LLM server with deployment, monitoring, injection and auto-scaling service:

enova pilot run --model mistralai/Mistral-7B-Instruct-v0.1

Use proxy to download LLMs:

enova pilot run --model mistralai/Mistral-7B-Instruct-v0.1 --hf_proxy xxx

Note

First, we use docker-compose to start grafana, prometheus, opentelmentry-collector, tempo, dcgm modules.
Second, we start a streamlit webui service.
Third, we launch the llm server with default vllm backend.

Tip

The default port of LLM service is 9199.
The default port of grafana server is 32827.
The default port of LLM webUI server is 8501.
The default port of ENOVA application server is 8182.

Check Deployed LLM service via ENOVA Application Server:

http://localhost:8182/instance

Test the Deployed LLM service with an prompt:

Use WebUI:

http://localhost:8501

Use Shell:

curl -X POST http://localhost:9199/generate \
-d '{
"prompt": "San Francisco is a",
"max_tokens": 1024,
"temperature": 0.9,
"top_p": 0.9
}'

Monitor the LLM Service Quality via ENOVA Application Server:

http://localhost:8182/instance

Stop all service

enova pilot stop --service all

🏠 LLM Deployment Service

In addition to offering an all-in-one solution for service deployment, monitoring, and autoscaling, ENOVA also provides support for single modules.

The LLM deployment service facilitates the deployment of LLMs and provides a stable API for accessing LLMs.

enova enode run --model mistralai/Mistral-7B-Instruct-v0.1

Note

The LLM server is launched with default vllm backend. OpenAI API and Generate API are both supported. vllm config can be specified using command-line parameters like:

enova enode run --model mistralai/Mistral-7B-Instruct-v0.1 --host 127.0.0.1 --port 9199

LLM WebUI Service

This service features a WebUI page for dialog interaction, where the serving host and port for the LLM server, and the host and port for the WebUI service, are configurable parameters.

enova webui run --serving_host 127.0.0.1 --serving_port 9199 --host 127.0.0.1 --port 8501

Autoscaling Module

The autoscaling service is automatically launched and managed by escaler module.

Request Injection Module

We implemented a request injection module using JMeter to simulate real user requests for evaluating LLM performance. The module allows simulation of request arrival probabilities using two modes: Poisson distribution and normal distribution. Further details on the injection operation are available at:

http://localhost:8182/instance

🏠 LLM Monitoring Service

Manage LLM Monitoring Service

The monitoring system is designed for monitoring and autoscaling, which contains real-time data collection, storage, and consumption. We can manage the LLM monitoring service via:

Start llm monitoring service

enova mon start

Check service status

enova mon status

Stop monitoring service

enova mon stop

Monitoring Metrics

Monitoring metrics are collected using the DCGM exporter, Prometheus exporters, and the OpenTelemetry collector. A brief description is provided in the following tables. For more details, please refer to the Grafana dashboard.

http://localhost:32827/dashboards

# default user / passwd
admin / grafana

Metric Type	Metric Description
API Service	The number of requests sent to LLM services per second
API Service	The number of requests processed by LLM services per second
API Service	The number of requests successfully processed per second
API Service	The success rate of requests processed by LLM services per second
API Service	The number of requests being processed by LLM services
API Service	The average execution time per request processed by LLM services
API Service	The average request size of requests per second
API Service	The average response size of requests per second
LLM Performance	The average prompt throughput per second
LLM Performance	The average generation throughput per second
LLM Performance	The number of requests being processed by the deployed LLM
LLM Performance	The number of requests being pended by the deployed LLM
LLM Performance	The utilization ratio of memory allocated for KV cache
GPU Utilization	DCGM Metrics, like DCGM_FI_DEV_GPU_UTIL.

📚 Reference

@inproceedings{tao2024ENOVA,
  title={ENOVA: Autoscaling towards Cost-effective and Stable Serverless LLM Serving},
  author={Tao Huang and Pengfei Chen and Kyoka Gong and Jocky Hawk and Zachary Bright and Wenxin Xie and Kecheng Huang and Zhi Ji},
  booktitle={arXiv preprint arXiv:},
  year={2024}
}

🤝 Join Our Community

We use Slack workspace for the collaboration on building ENOVA.

Slack workspace

vamoko / ENOVA