Skip to content

Latest commit

 

History

History
169 lines (106 loc) · 6.06 KB

Readme.md

File metadata and controls

169 lines (106 loc) · 6.06 KB
Raw

TorchServe Llama Chatapp

This is an example showing how to deploy a Llama chat app using TorchServe. We use streamlit to create the app

We are using llama-cpp-python in this example

You can run this example on your laptop to understand how to use TorchServe

Quick Start Guide

To get started with TorchServe, you need to run the following

# 1: Set HF Token as Env variable
export HUGGINGFACE_TOKEN=<Token> # get this from your HuggingFace account

# 2: Build TorchServe Chat Bot Image for Serving
./examples/LLM/llama/chat_app/docker/build_image.sh

# 3: Launch the streamlit app for server & client
docker run --rm -it --platform linux/amd64 -p 127.0.0.1:8080:8080 -p 127.0.0.1:8081:8081 -p 127.0.0.1:8082:8082 -p 127.0.0.1:8084:8084 -p 127.0.0.1:8085:8085 -v <model-store>:/home/model-server/model-store -e MODEL_NAME=meta-llama/Meta-Llama-3-8B-Instruct pytorch/torchserve:chat_bot

In step 3, <model-store> is a location where you want the model to be downloaded

What to expect

This launches two streamlit apps

  1. TorchServe Server app to start/stop TorchServe, load model, scale up/down workers, configure dynamic batch_size ( Currently llama-cpp-python doesn't support batch_size > 1)
    • Since this app is targeted for Apple M1/M2 laptops, we load a 4-bit quantized version of llama using llama-cpp-python.
  2. Client chat app where you can chat with the model . There is a slider to send concurrent requests to the model. The current app doesn't have a good mechanism to show multiple responses in parallel. You can notice streaming response for the first request followed by a complete response for the next request.

Currently, this launches Meta-Llama-3-8B-Instruct with 4-bit quantization running on CPU.

To make use of M1/M2 GPU, you can follow the below guide to do a standalone TorchServe installation.

Architecture

Chatbot Architecture

Pre-requisites

The following example has been tested on M1 Mac. Before you install TorchServe, make sure you have the following installed

  1. JDK 17

Make sure your javac version is 17.x.x

javac --version
javac 17.0.8

You can download it from java 2) Install conda with support for arm64

  1. Since we are running this example on Mac, we will use the Meta-Llama-3-8B-Instruct model. Download Meta-Llama-3-8B-Instruct weights by following instructions here

  2. Install streamlit with

python -m pip install -r requirements.txt

Steps

Install TorchServe

Install TorchServe with the following steps

python ts_scripts/install_dependencies.py
pip install torchserve torch-model-archiver torch-workflow-archiver

Package model for TorchServe

Run this script to create llamacpp.tar.gz to be loaded in TorchServe

source package_llama.sh <path to llama snapshot folder>

This creates the quantized weights in $LLAMA_WEIGHTS

For subsequent runs, we don't need to regenerate these weights. We only need to package the handler, model-config.yaml in the tar file.

Hence, you can skip the model generation by running the script as follows

source package_llama.sh <path to llama2 snapshot folder>  false

You might need to run the below command if the script output indicates it.

sudo xcodebuild -license

The script is setting an env variable LLAMA_Q4_MODEL and using this in the handler. In an actual use-case, you would set the path to the weights in model-config.yaml

handler:
    model_name: "llama-cpp"
    model_path: "<absolute path to the weights file"

Start TorchServe

We launch a streamlit app to configure TorchServe. This opens a UI in your browser, which you can use to start/stop TorchServe, register model, change some of the TorchServe parameters

streamlit run torchserve_server_app.py

You can check the model status on the app to make sure the model is ready to receive requests

Server

Client Chat App

We launch a streamlit app from which a client can send requests to TorchServe. The reference app used is here

streamlit run client_app.py

You can change the model parameters and ask the server questions in the following format

Question: What is the closest star to Earth ? Answer:

results in

Question: What is the closest star to Earth ? Answer: The closest star to Earth is Proxima Centauri, which is located about 4. nobody knows if there is other life out there similar to ours or not, but it's pretty cool that we know of a star so close to us!

Client

Experiments

You can launch a second client app from another terminal.

You can send requests simultaneously to see how quickly TorchServe responds

Dynamic Batching

Batching requests is a strategy used to make efficient use of compute resources. TorchServe supports dynamic batching, where the frontend can batch requests if the requests arrive within the max_batch_delay time. You can make use of dynamic batching in TorchServe by configuring the batch_size and max_batch_delay parameters in TorchServe. You can do this on the Server app.

You can read more about batching in TorchServe here

Batch Size

Backend Workers

TorchServe's backend workers perform the actual model inference. You can increase the number of backend workers in TorchServe by configuring min_workers parameter in TorchServe. You can do this on the Server app.

The number of workers can be autoscaled based on the traffic and usage patterns.

Number of workers = Number of TorchServe python processes

So to reduce latency and improve performance, we can increase the number of workers for parallel processing of requests.

Workers