The hosted Aviary Explorer is not available anymore. Visit Anyscale to experience models served with RayLLM.
RayLLM (formerly known as Aviary) is an LLM serving solution that makes it easy to deploy and manage a variety of open source LLMs, built on Ray Serve. It does this by:
- Providing an extensive suite of pre-configured open source LLMs, with defaults that work out of the box.
- Supporting Transformer models hosted on Hugging Face Hub or present on local disk.
- Simplifying the deployment of multiple LLMs
- Simplifying the addition of new LLMs
- Offering unique autoscaling support, including scale-to-zero.
- Fully supporting multi-GPU & multi-node model deployments.
- Offering high performance features like continuous batching, quantization and streaming.
- Providing a REST API that is similar to OpenAI's to make it easy to migrate and cross test them.
- Supporting multiple LLM backends out of the box, including vLLM and TensorRT-LLM.
In addition to LLM serving, it also includes a CLI and a web frontend (Aviary Explorer) that you can use to compare the outputs of different models directly, rank them by quality, get a cost and latency estimate, and more.
RayLLM supports continuous batching and quantization by integrating with vLLM. Continuous batching allows you to get much better throughput and latency than static batching. Quantization allows you to deploy compressed models with cheaper hardware requirements and lower inference costs. See quantization guide for more details on running quantized models on RayLLM.
RayLLM leverages Ray Serve, which has native support for autoscaling and multi-node deployments. RayLLM can scale to zero and create new model replicas (each composed of multiple GPU workers) in response to demand.
The guide below walks you through the steps required for deployment of RayLLM on Ray Serve.
We highly recommend using the official anyscale/ray-llm Docker image to run RayLLM. Manually installing RayLLM is currently not a supported use-case due to specific dependencies required, some of which are not available on pip.
cache_dir=${XDG_CACHE_HOME:-$HOME/.cache}
docker run -it --gpus all --shm-size 1g -p 8000:8000 -e HF_HOME=~/data -v $cache_dir:~/data anyscale/ray-llm:latest bash
# Inside docker container
serve run ~/serve_configs/amazon--LightGPT.yamlRayLLM uses Ray Serve, so it can be deployed on Ray Clusters.
Currently, we only have a guide and pre-configured YAML file for AWS deployments. Make sure you have exported your AWS credentials locally.
export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...
export AWS_SESSION_TOKEN=...Start by cloning this repo to your local machine.
You may need to specify your AWS private key in the deploy/ray/rayllm-cluster.yaml file.
See Ray on Cloud VMs page in
Ray documentation for more details.
git clone https://github.com/ray-project/ray-llm.git
cd ray-llm
# Start a Ray Cluster (This will take a few minutes to start-up)
ray up deploy/ray/rayllm-cluster.yaml# Connect to the Head node of your Ray Cluster (This will take several minutes to autoscale)
ray attach deploy/ray/rayllm-cluster.yaml
# Deploy the LightGPT model.
serve run serve_configs/amazon--LightGPT.yamlYou can deploy any model in the models directory of this repo,
or define your own model YAML file and run that instead.
For Kubernetes deployments, please see our documentation for deploying on KubeRay.
Once the models are deployed, you can install a client outside of the Docker container to query the backend.
pip install "rayllm @ git+https://github.com/ray-project/ray-llm.git"You can query your RayLLM deployment in many ways.
In all cases start out by doing:
export ENDPOINT_URL="http://localhost:8000/v1"This is because your deployment is running locally, but you can also access remote deployments (in which case you would set ENDPOINT_URL to a remote URL).
You can use curl at the command line to query your deployed LLM:
% curl $ENDPOINT_URL/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "meta-llama/Llama-2-7b-chat-hf",
"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Hello!"}],
"temperature": 0.7
}'{
"id":"meta-llama/Llama-2-7b-chat-hf-308fc81f-746e-4682-af70-05d35b2ee17d",
"object":"text_completion","created":1694809775,
"model":"meta-llama/Llama-2-7b-chat-hf",
"choices":[
{
"message":
{
"role":"assistant",
"content":"Hello there! *adjusts glasses* It's a pleasure to meet you! Is there anything I can help you with today? Have you got a question or a task you'd like me to assist you with? Just let me know!"
},
"index":0,
"finish_reason":"stop"
}
],
"usage":{"prompt_tokens":30,"completion_tokens":53,"total_tokens":83}}
Use the requests library to connect with Python. Use this script to receive a streamed response, automatically parse the outputs, and print just the content.
import os
import json
import requests
s = requests.Session()
api_base = os.getenv("ENDPOINT_URL")
url = f"{api_base}/chat/completions"
body = {
"model": "meta-llama/Llama-2-7b-chat-hf",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Tell me a long story with many words."}
],
"temperature": 0.7,
"stream": True,
}
with s.post(url, json=body, stream=True) as response:
for chunk in response.iter_lines(decode_unicode=True):
if chunk is not None:
try:
# Get data from reponse chunk
chunk_data = chunk.split("data: ")[1]
# Get message choices from data
choices = json.loads(chunk_data)["choices"]
# Pick content from first choice
content = choices[0]["delta"]["content"]
print(content, end="", flush=True)
except json.decoder.JSONDecodeError:
# Chunk was not formatted as expected
pass
except KeyError:
# No message was contained in the chunk
pass
print("")RayLLM uses an OpenAI-compatible API, allowing us to use the OpenAI
SDK to access our deployments. To do so, we need to set the OPENAI_API_BASE env var.
export OPENAI_API_BASE=http://localhost:8000/v1
export OPENAI_API_KEY='not_a_real_key'import openai
# List all models.
models = openai.Model.list()
print(models)
# Note: not all arguments are currently supported and will be ignored by the backend.
chat_completion = openai.ChatCompletion.create(
model="meta-llama/Llama-2-7b-chat-hf",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Say 'test'."}
],
temperature=0.7
)
print(chat_completion)To install RayLLM and its dependencies, run the following command:
pip install "rayllm @ git+https://github.com/ray-project/ray-llm.git"RayLLM consists of a set of configurations and utilities for deploying LLMs on Ray Serve, in addition to a frontend (Aviary Explorer), both of which come with additional dependencies. To install the dependencies for the frontend run the following commands:
pip install "rayllm[frontend] @ git+https://github.com/ray-project/ray-llm.git"The backend dependencies are heavy weight, and quite large. We recommend using the official
anyscale/ray-llm image. Installing the backend manually is not a supported usecase.
Ray collects basic, non-identifiable usage statistics to help us improve the project. For more information on what is collected and how to opt-out, see the Usage Stats Collection page in Ray documentation.
RayLLM uses the Ray Serve CLI that allows you to interact with deployed models.
# Start a new model in Ray Serve from provided configuration
serve run serve_configs/<model_config_path>
# Get the status of the running deployments
serve status
# Get the current config of current live Serve applications
serve config
# Shutdown all Serve applications
serve shutdownYou can easily add new models by adding two configuration files. To learn more about how to customize or add new models, see the Model Registry.
The easiest way is to copy the configuration of the existing model's YAML file and modify it. See models/README.md for more details.
Run multiple models at once by aggregating the Serve configs for different models into a single, unified config. For example, use this config to run the LightGPT and Llama-2-7b-chat model in a single Serve application:
# File name: serve_configs/config.yaml
applications:
- name: router
import_path: rayllm.backend:router_application
route_prefix: /
args:
models:
- ./models/continuous_batching/amazon--LightGPT.yaml
- ./models/continuous_batching/meta-llama--Llama-2-7b-chat-hf.yamlThe config includes both models in the model argument for the router. Additionally, the Serve configs for both model applications are included. Save this unified config file to the serve_configs/ folder.
Run the config to deploy the models:
serve run serve_configs/<config.yaml>All our default model configurations enforce a model to be deployed on one node for high performance. However, you can easily change this if you want to deploy a model across nodes for lower cost or GPU availability. In order to do that, go to the YAML file in the model registry and change placement_strategy to PACK instead of STRICT_PACK.
There can be several reasons for the deployment not starting or not working correctly. Here are some things to check:
- You might have specified an invalid model id.
- Your model may require resources that are not available on the cluster. A common issue is that the model requires Ray custom resources (eg.
accelerator_type_a10) in order to be scheduled on the right node type, while your cluster is missing those custom resources. You can either modify the model configuration to remove those custom resources or better yet, add them to the node configuration of your Ray cluster. You can debug this issue by looking at Ray Autoscaler logs (monitor.log). - Your model is a gated Hugging Face model (eg. meta-llama). In that case, you need to set the
HUGGING_FACE_HUB_TOKENenvironment variable cluster-wide. You can do that either in the Ray cluster configuration or by setting it before runningserve run - Your model may be running out of memory. You can usually spot this issue by looking for keywords related to "CUDA", "memory" and "NCCL" in the replica logs or
serve runoutput. In that case, consider reducing themax_batch_prefill_tokensandmax_batch_total_tokens(if applicable). See models/README.md for more information on those parameters.
In general, Ray Dashboard is a useful debugging tool, letting you monitor your Ray Serve / LLM application and access Ray logs.
A good sanity check is deploying the test model in tests/models/. If that works, you know you can deploy a model.
The OpenAI create() commands allow you to specify the API_KEY and API_BASE. So you can do something like this.
# Call your self-hosted model running on the local host:
OpenAI.ChatCompletion.create(api_base="http://localhost:8000/v1", api_key="",...)
# Call OpenAI. Set OPENAI_API_KEY to your key and unset OPENAI_API_BASE
OpenAI.ChatCompletion.create(api_key="OPENAI_API_KEY", ...)We are eager to help you get started with RayLLM. You can get help on:
For bugs or for feature requests, please submit them here.
We are also interested in accepting contributions. Those could be anything from a new evaluator, to integrating a new model with a yaml file, to more. Feel free to post an issue first to get our feedback on a proposal first, or just file a PR and we commit to giving you prompt feedback.
We use pre-commit hooks to ensure that all code is formatted correctly.
Make sure to pip install pre-commit and then run pre-commit install.
You can also run ./format to run the hooks manually.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent