TorchServe quickstart chatbot example (#3003)

* TorchServe quickstart chatbot example

* Added more details in Readme

* lint failure

* code cleanup

* review comments

---------

Co-authored-by: Mark Saroufim <[email protected]>

Author: agunapal

examples/LLM/llama2/chat_app/Readme.md

@@ -9,6 +9,33 @@ We are using [llama-cpp-python](https://github.com/abetlen/llama-cpp-python) in
 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 Image for Serving llama2-7b model with 4-bit quantization
+./examples/llm/llama2/chat_app/docker/build_image.sh meta-llama/Llama-2-7b-chat-hf
+
+# 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 pytorch/torchserve:meta-llama---Llama-2-7b-chat-hf
+```
+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 llama2 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 llama2-7b model 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](./screenshots/architecture.png)

examples/LLM/llama2/chat_app/docker/Dockerfile

@@ -0,0 +1,26 @@
+ARG BASE_IMAGE=pytorch/torchserve:latest-gpu
+
+FROM $BASE_IMAGE as server
+ARG BASE_IMAGE
+ARG EXAMPLE_DIR
+ARG MODEL_NAME
+ARG HUGGINGFACE_TOKEN
+
+USER root
+
+ENV MODEL_NAME=$MODEL_NAME
+
+RUN --mount=type=cache,id=apt-dev,target=/var/cache/apt \
+    apt-get update && \
+    apt-get install libopenmpi-dev git -y
+
+COPY $EXAMPLE_DIR/requirements.txt /home/model-server/chat_bot/requirements.txt
+RUN pip install -r /home/model-server/chat_bot/requirements.txt && huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+COPY $EXAMPLE_DIR  /home/model-server/chat_bot
+COPY $EXAMPLE_DIR/dockerd-entrypoint.sh /usr/local/bin/dockerd-entrypoint.sh
+COPY $EXAMPLE_DIR/config.properties /home/model-server/config.properties
+
+WORKDIR /home/model-server/chat_bot
+RUN chmod +x /usr/local/bin/dockerd-entrypoint.sh \
+    && chown -R model-server /home/model-server

examples/LLM/llama2/chat_app/docker/build_image.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+# Check if there are enough arguments
+if [ "$#" -eq 0 ] || [ "$#" -gt 1 ]; then
+  echo "Usage: $0 <HF Model>"
+  exit 1
+fi
+
+MODEL_NAME=$(echo "$1" | sed 's/\//---/g')
+echo "Model: " $MODEL_NAME
+
+BASE_IMAGE="pytorch/torchserve:latest-cpu"
+
+DOCKER_TAG="pytorch/torchserve:${MODEL_NAME}"
+
+# Get relative path of example dir
+EXAMPLE_DIR=$(dirname "$(readlink -f "$0")")
+ROOT_DIR=${EXAMPLE_DIR}/../../../../..
+ROOT_DIR=$(realpath "$ROOT_DIR")
+EXAMPLE_DIR=$(echo "$EXAMPLE_DIR" | sed "s|$ROOT_DIR|./|")
+
+# Build docker image for the application
+DOCKER_BUILDKIT=1 docker buildx build --platform=linux/amd64 --file ${EXAMPLE_DIR}/Dockerfile --build-arg BASE_IMAGE="${BASE_IMAGE}" --build-arg EXAMPLE_DIR="${EXAMPLE_DIR}" --build-arg MODEL_NAME="${MODEL_NAME}"  --build-arg HUGGINGFACE_TOKEN -t "${DOCKER_TAG}" .
+
+echo "Run the following command to start the chat bot"
+echo ""
+echo 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 $(pwd)/model_store_1:/home/model-server/model-store $DOCKER_TAG
+echo ""

examples/LLM/llama2/chat_app/docker/client_app.py

@@ -0,0 +1,118 @@
+import json
+import os
+from concurrent.futures import ThreadPoolExecutor
+
+import requests
+import streamlit as st
+
+MODEL_NAME = os.environ["MODEL_NAME"]
+
+# App title
+st.set_page_config(page_title="TorchServe Chatbot")
+
+with st.sidebar:
+    st.title("TorchServe Chatbot")
+
+    st.session_state.model_loaded = False
+    try:
+        res = requests.get(url="http://localhost:8080/ping")
+        res = requests.get(url=f"http://localhost:8081/models/{MODEL_NAME}")
+        status = "NOT READY"
+        if res.status_code == 200:
+            status = json.loads(res.text)[0]["workers"][0]["status"]
+
+        if status == "READY":
+            st.session_state.model_loaded = True
+            st.success("Proceed to entering your prompt message!", icon="👉")
+        else:
+            st.warning("Model not loaded in TorchServe", icon="⚠️")
+
+    except requests.ConnectionError:
+        st.warning("TorchServe is not up. Try again", icon="⚠️")
+
+    if st.session_state.model_loaded:
+        st.success(f"Model loaded: {MODEL_NAME}!", icon="👉")
+
+    st.subheader("Model parameters")
+    temperature = st.sidebar.slider(
+        "temperature", min_value=0.1, max_value=1.0, value=0.5, step=0.1
+    )
+    top_p = st.sidebar.slider(
+        "top_p", min_value=0.1, max_value=1.0, value=0.5, step=0.1
+    )
+    max_new_tokens = st.sidebar.slider(
+        "max_new_tokens", min_value=48, max_value=512, value=50, step=4
+    )
+    concurrent_requests = st.sidebar.select_slider(
+        "concurrent_requests", options=[2**j for j in range(0, 8)]
+    )
+
+# Store LLM generated responses
+if "messages" not in st.session_state.keys():
+    st.session_state.messages = [
+        {"role": "assistant", "content": "How may I assist you today?"}
+    ]
+
+# Display or clear chat messages
+for message in st.session_state.messages:
+    with st.chat_message(message["role"]):
+        st.write(message["content"])
+
+
+def clear_chat_history():
+    st.session_state.messages = [
+        {"role": "assistant", "content": "How may I assist you today?"}
+    ]
+
+
+st.sidebar.button("Clear Chat History", on_click=clear_chat_history)
+
+
+def generate_model_response(prompt_input, executor):
+    string_dialogue = (
+        "Question: What are the names of the planets in the solar system? Answer: "
+    )
+    headers = {"Content-type": "application/json", "Accept": "text/plain"}
+    url = f"http://127.0.0.1:8080/predictions/{MODEL_NAME}"
+    data = json.dumps(
+        {
+            "prompt": prompt_input,
+            "params": {
+                "max_new_tokens": max_new_tokens,
+                "top_p": top_p,
+                "temperature": temperature,
+            },
+        }
+    )
+    res = [
+        executor.submit(requests.post, url=url, data=data, headers=headers, stream=True)
+        for i in range(concurrent_requests)
+    ]
+
+    return res, max_new_tokens
+
+
+# User-provided prompt
+if prompt := st.chat_input():
+    st.session_state.messages.append({"role": "user", "content": prompt})
+    with st.chat_message("user"):
+        st.write(prompt)
+
+# Generate a new response if last message is not from assistant
+if st.session_state.messages[-1]["role"] != "assistant":
+    with st.chat_message("assistant"):
+        with st.spinner("Thinking..."):
+            with ThreadPoolExecutor() as executor:
+                futures, max_tokens = generate_model_response(prompt, executor)
+                placeholder = st.empty()
+                full_response = ""
+                count = 0
+                for future in futures:
+                    response = future.result()
+                    for chunk in response.iter_content(chunk_size=None):
+                        if chunk:
+                            data = chunk.decode("utf-8")
+                            full_response += data
+                            placeholder.markdown(full_response)
+    message = {"role": "assistant", "content": full_response}
+    st.session_state.messages.append(message)

examples/LLM/llama2/chat_app/docker/config.properties

@@ -0,0 +1,9 @@
+metrics_mode=prometheus
+model_metrics_auto_detect=true
+inference_address=http://0.0.0.0:8080
+management_address=http://0.0.0.0:8081
+metrics_address=http://0.0.0.0:8082
+number_of_netty_threads=32
+job_queue_size=1000
+model_store=/home/model-server/model-store
+workflow_store=/home/model-server/wf-store

examples/LLM/llama2/chat_app/docker/dockerd-entrypoint.sh

@@ -0,0 +1,81 @@
+#!/bin/bash
+set -e
+
+export LLAMA2_Q4_MODEL=/home/model-server/model-store/$MODEL_NAME/model/ggml-model-q4_0.gguf
+
+
+create_model_cfg_yaml() {
+  # Define the YAML content with a placeholder for the model name
+  yaml_content="# TorchServe frontend parameters\nminWorkers: 1\nmaxWorkers: 1\nresponseTimeout: 1200\n#deviceType: \"gpu\"\n#deviceIds: [0,1]\n#torchrun:\n#    nproc-per-node: 1\n\nhandler:\n    model_name: \"${2}\"\n    manual_seed: 40"
+
+  # Create the YAML file with the specified model name
+  echo -e "$yaml_content" > "model-config-${1}.yaml"
+}
+
+create_model_archive() {
+    MODEL_NAME=$1
+    MODEL_CFG=$2
+    echo "Create model archive for ${MODEL_NAME} if it doesn't already exist"
+    if [ -d "/home/model-server/model-store/$MODEL_NAME" ]; then
+        echo "Model archive for $MODEL_NAME exists."
+    fi
+    if [ -d "/home/model-server/model-store/$MODEL_NAME/model" ]; then
+        echo "Model already download"
+        mv /home/model-server/model-store/$MODEL_NAME/model /home/model-server/model-store/
+    else
+        echo "Model needs to be downloaded"
+    fi
+    torch-model-archiver --model-name "$MODEL_NAME" --version 1.0 --handler llama_cpp_handler.py --config-file $MODEL_CFG -r requirements.txt --archive-format no-archive --export-path /home/model-server/model-store -f
+    if [ -d "/home/model-server/model-store/model" ]; then
+        mv /home/model-server/model-store/model /home/model-server/model-store/$MODEL_NAME/
+    fi
+}
+
+download_model() {
+   MODEL_NAME=$1
+   HF_MODEL_NAME=$2
+    if [ -d "/home/model-server/model-store/$MODEL_NAME/model" ]; then
+        echo "Model $HF_MODEL_NAME already downloaded"
+    else
+        echo "Downloading  model $HF_MODEL_NAME"
+        python Download_model.py --model_path /home/model-server/model-store/$MODEL_NAME/model --model_name $HF_MODEL_NAME
+    fi
+}
+
+quantize_model() {
+    if [ ! -f "$LLAMA2_Q4_MODEL" ]; then
+        tmp_model_name=$(echo "$MODEL_NAME" | sed 's/---/--/g')
+        directory_path=/home/model-server/model-store/$MODEL_NAME/model/models--$tmp_model_name/snapshots/
+        HF_MODEL_SNAPSHOT=$(find $directory_path -type d -mindepth 1)
+        echo "Cleaning up previous build of llama-cpp"
+        git clone https://github.com/ggerganov/llama.cpp.git build
+        cd build
+        make
+        python -m pip install -r requirements.txt
+
+        echo "Convert the 7B model to ggml FP16 format"
+        python convert.py $HF_MODEL_SNAPSHOT --outfile ggml-model-f16.gguf
+
+        echo "Quantize the model to 4-bits (using q4_0 method)"
+        ./quantize ggml-model-f16.gguf $LLAMA2_Q4_MODEL q4_0
+
+        cd ..
+        echo "Saved quantized model weights to $LLAMA2_Q4_MODEL"
+    fi
+}
+
+HF_MODEL_NAME=$(echo "$MODEL_NAME" | sed 's/---/\//g')
+if [[ "$1" = "serve" ]]; then
+    shift 1
+    create_model_cfg_yaml $MODEL_NAME $HF_MODEL_NAME
+    create_model_archive $MODEL_NAME "model-config-$MODEL_NAME.yaml"
+    download_model $MODEL_NAME $HF_MODEL_NAME
+    quantize_model
+    streamlit run torchserve_server_app.py --server.port 8084 &
+    streamlit run client_app.py --server.port 8085
+else
+    eval "$@"
+fi
+
+# prevent docker exit
+tail -f /dev/null

examples/LLM/llama2/chat_app/docker/llama_cpp_handler.py

@@ -0,0 +1,67 @@
+import logging
+import os
+from abc import ABC
+
+import torch
+from llama_cpp import Llama
+
+from ts.protocol.otf_message_handler import send_intermediate_predict_response
+from ts.torch_handler.base_handler import BaseHandler
+
+logger = logging.getLogger(__name__)
+
+
+class LlamaCppHandler(BaseHandler, ABC):
+    def __init__(self):
+        super(LlamaCppHandler, self).__init__()
+        self.initialized = False
+
+    def initialize(self, ctx):
+        """In this initialize function, the HF large model is loaded and
+        partitioned using DeepSpeed.
+        Args:
+            ctx (context): It is a JSON Object containing information
+            pertaining to the model artifacts parameters.
+        """
+        model_path = os.environ["LLAMA2_Q4_MODEL"]
+        model_name = ctx.model_yaml_config["handler"]["model_name"]
+        seed = int(ctx.model_yaml_config["handler"]["manual_seed"])
+        torch.manual_seed(seed)
+
+        self.model = Llama(model_path=model_path)
+        logger.info(f"Loaded {model_name} model successfully")
+
+    def preprocess(self, data):
+        assert (
+            len(data) == 1
+        ), "llama-cpp-python is currently only supported with batch_size=1"
+        for row in data:
+            item = row.get("body")
+            return item
+
+    def inference(self, data):
+        params = data["params"]
+        tokens = self.model.tokenize(bytes(data["prompt"], "utf-8"))
+        generation_kwargs = dict(
+            tokens=tokens,
+            temp=params["temperature"],
+            top_p=params["top_p"],
+        )
+        count = 0
+        for token in self.model.generate(**generation_kwargs):
+            if count >= params["max_new_tokens"]:
+                break
+
+            count += 1
+            new_text = self.model.detokenize([token])
+            send_intermediate_predict_response(
+                [new_text],
+                self.context.request_ids,
+                "Intermediate Prediction success",
+                200,
+                self.context,
+            )
+        return [""]
+
+    def postprocess(self, output):
+        return output

examples/LLM/llama2/chat_app/docker/requirements.txt

@@ -0,0 +1,4 @@
+transformers
+llama-cpp-python
+streamlit>=1.26.0
+requests_futures