Skip to content

Latest commit

 

History

History
431 lines (333 loc) · 15.9 KB

fastchat_quickstart.md

File metadata and controls

431 lines (333 loc) · 15.9 KB

Serving using IPEX-LLM and FastChat

FastChat is an open platform for training, serving, and evaluating large language model based chatbots. You can find the detailed information at their homepage.

IPEX-LLM can be easily integrated into FastChat so that user can use IPEX-LLM as a serving backend in the deployment.

Table of Contents

Quick Start

This quickstart guide walks you through installing and running FastChat with ipex-llm.

1. Install IPEX-LLM with FastChat

To run on CPU, you can install ipex-llm as follows:

pip install --pre --upgrade ipex-llm[serving,all]

To add GPU support for FastChat, you may install ipex-llm as follows:

pip install --pre --upgrade ipex-llm[xpu,serving] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/

2. Start the service

Launch controller

You need first run the fastchat controller

python3 -m fastchat.serve.controller

If the controller run successfully, you can see the output like this:

Uvicorn running on http://localhost:21001

Launch model worker(s) and load models

Using IPEX-LLM in FastChat does not impose any new limitations on model usage. Therefore, all Hugging Face Transformer models can be utilized in FastChat.

IPEX-LLM worker

To integrate IPEX-LLM with FastChat efficiently, we have provided a new model_worker implementation named ipex_llm_worker.py.

# On CPU
# Available low_bit format including sym_int4, sym_int8, bf16 etc.
python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "cpu"

# On GPU
# Available low_bit format including sym_int4, sym_int8, fp16 etc.
source /opt/intel/oneapi/setvars.sh
export USE_XETLA=OFF
# [optional] under most circumstances, the following environment variable may improve performance, but sometimes this may also cause performance degradation
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1

python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "xpu"

We have also provided an option --load-low-bit-model to load models that have been converted and saved into disk using the save_low_bit interface as introduced in this document.

Check the following examples:

# Or --device "cpu"
python -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path /Low/Bit/Model/Path --trust-remote-code --device "xpu" --load-low-bit-model

For self-speculative decoding example:

You can use IPEX-LLM to run self-speculative decoding example. Refer to here for more details on intel MAX GPUs. Refer to here for more details on intel CPUs.

# Available low_bit format only including bf16 on CPU.
source ipex-llm-init -t
python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path lmsys/vicuna-7b-v1.5 --low-bit "bf16" --trust-remote-code --device "cpu" --speculative

# Available low_bit format only including fp16 on GPU.
source /opt/intel/oneapi/setvars.sh
export ENABLE_SDP_FUSION=1
export SYCL_CACHE_PERSISTENT=1
# [optional] under most circumstances, the following environment variable may improve performance, but sometimes this may also cause performance degradation
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path lmsys/vicuna-7b-v1.5 --low-bit "fp16" --trust-remote-code --device "xpu" --speculative

You can get output like this:

2024-04-12 18:18:09 | INFO | ipex_llm.transformers.utils | Converting the current model to sym_int4 format......
2024-04-12 18:18:11 | INFO | model_worker | Register to controller
2024-04-12 18:18:11 | ERROR | stderr | INFO:     Started server process [126133]
2024-04-12 18:18:11 | ERROR | stderr | INFO:     Waiting for application startup.
2024-04-12 18:18:11 | ERROR | stderr | INFO:     Application startup complete.
2024-04-12 18:18:11 | ERROR | stderr | INFO:     Uvicorn running on http://localhost:21002

For a full list of accepted arguments, you can refer to the main method of the ipex_llm_worker.py

IPEX-LLM vLLM worker

We also provide the vllm_worker which uses the vLLM engine (on CPU / GPU) for better hardware utilization.

To run using the vLLM_worker, we don't need to change model name, just simply uses the following command:

# On CPU
python3 -m ipex_llm.serving.fastchat.vllm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --device cpu

# On GPU
source /opt/intel/oneapi/setvars.sh
export USE_XETLA=OFF
# [optional] under most circumstances, the following environment variable may improve performance, but sometimes this may also cause performance degradation
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
python3 -m ipex_llm.serving.fastchat.vllm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --device xpu --load-in-low-bit "sym_int4" --enforce-eager

Note

The environment variable SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS determines the usage of immediate command lists for task submission to the GPU. While this mode typically enhances performance, exceptions may occur. Please consider experimenting with and without this environment variable for best performance. For more details, you can refer to this article.

Launch multiple workers

Sometimes we may want to start multiple workers for the best performance. For running in CPU, you may want to seperate multiple workers in different sockets. Assuming each socket have 48 physicall cores, then you may want to start two workers using the following example:

export OMP_NUM_THREADS=48
numactl -C 0-47 -m 0 python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "cpu" &

# All the workers other than the first worker need to specify a different worker port and corresponding worker-address
numactl -C 48-95 -m 1 python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "cpu" --port 21003 --worker-address "http://localhost:21003" &

For GPU, we may want to start two workers using different GPUs. To achieve this, you should use ZE_AFFINITY_MASK environment variable to select different GPUs for different workers. Below shows an example:

ZE_AFFINITY_MASK=1 python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "xpu" &

# All the workers other than the first worker need to specify a different worker port and corresponding worker-address
ZE_AFFINITY_MASK=2 python3 -m ipex_llm.serving.fastchat.ipex_llm_worker --model-path REPO_ID_OR_YOUR_MODEL_PATH --low-bit "sym_int4" --trust-remote-code --device "xpu" --port 21003 --worker-address "http://localhost:21003" &

If you are not sure the effect of ZE_AFFINITY_MASK, then you could set ZE_AFFINITY_MASK and check the result of sycl-ls.

Launch Gradio web server

When you have started the controller and the worker, you can start web server as follows:

python3 -m fastchat.serve.gradio_web_server

This is the user interface that users will interact with.

By following these steps, you will be able to serve your models using the web UI with IPEX-LLM as the backend. You can open your browser and chat with a model now.

Launch TGI Style API server

When you have started the controller and the worker, you can start TGI Style API server as follows:

python3 -m ipex_llm.serving.fastchat.tgi_api_server --host localhost --port 8000

You can use curl for observing the output of the api

Using /generate API

This is to send a sentence as inputs in the request, and is expected to receive a response containing model-generated answer.

curl -X POST -H "Content-Type: application/json" -d '{
  "inputs": "What is AI?",
  "parameters": {
    "best_of": 1,
    "decoder_input_details": true,
    "details": true,
    "do_sample": true,
    "frequency_penalty": 0.1,
    "grammar": {
      "type": "json",
      "value": "string"
    },
    "max_new_tokens": 32,
    "repetition_penalty": 1.03,
    "return_full_text": false,
    "seed": 0.1,
    "stop": [
      "photographer"
    ],
    "temperature": 0.5,
    "top_k": 10,
    "top_n_tokens": 5,
    "top_p": 0.95,
    "truncate": true,
    "typical_p": 0.95,
    "watermark": true
  }
}' http://localhost:8000/generate

Sample output:

{
    "details": {
        "best_of_sequences": [
            {
                "index": 0,
                "message": {
                    "role": "assistant",
                    "content": "\nArtificial Intelligence (AI) is a branch of computer science that attempts to simulate the way that the human brain works. It is a branch of computer "
                },
                "finish_reason": "length",
                "generated_text": "\nArtificial Intelligence (AI) is a branch of computer science that attempts to simulate the way that the human brain works. It is a branch of computer ",
                "generated_tokens": 31
            }
        ]
    },
    "generated_text": "\nArtificial Intelligence (AI) is a branch of computer science that attempts to simulate the way that the human brain works. It is a branch of computer ",
    "usage": {
        "prompt_tokens": 4,
        "total_tokens": 35,
        "completion_tokens": 31
    }
}

Using /generate_stream API

This is to send a sentence as inputs in the request, and a long connection will be opened to continuously receive multiple responses containing model-generated answer.

curl -X POST -H "Content-Type: application/json" -d '{
  "inputs": "What is AI?",
  "parameters": {
    "best_of": 1,
    "decoder_input_details": true,
    "details": true,
    "do_sample": true,
    "frequency_penalty": 0.1,
    "grammar": {
      "type": "json",
      "value": "string"
    },
    "max_new_tokens": 32,
    "repetition_penalty": 1.03,
    "return_full_text": false,
    "seed": 0.1,
    "stop": [
      "photographer"
    ],
    "temperature": 0.5,
    "top_k": 10,
    "top_n_tokens": 5,
    "top_p": 0.95,
    "truncate": true,
    "typical_p": 0.95,
    "watermark": true
  }
}' http://localhost:8000/generate_stream

Sample output:

data: {"token": {"id": 663359, "text": "", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 300560, "text": "\n", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 725120, "text": "Artificial Intelligence ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 734609, "text": "(AI) is ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 362235, "text": "a branch of computer ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 380983, "text": "science that attempts to ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 249979, "text": "simulate the way that ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 972663, "text": "the human brain ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 793301, "text": "works. It is a ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 501380, "text": "branch of computer ", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 673232, "text": "", "logprob": 0.0, "special": false}, "generated_text": null, "details": null, "special_ret": null}

data: {"token": {"id": 2, "text": "</s>", "logprob": 0.0, "special": true}, "generated_text": "\nArtificial Intelligence (AI) is a branch of computer science that attempts to simulate the way that the human brain works. It is a branch of computer ", "details": {"finish_reason": "eos_token", "generated_tokens": 31, "prefill_tokens": 4, "seed": 2023}, "special_ret": {"tensor": []}}

Launch RESTful API server

To start an OpenAI API server that provides compatible APIs using IPEX-LLM backend, you can launch the openai_api_server and follow this doc to use it.

When you have started the controller and the worker, you can start RESTful API server as follows:

python3 -m fastchat.serve.openai_api_server --host localhost --port 8000

You can use curl for observing the output of the api

You can format the output using jq

List Models

curl http://localhost:8000/v1/models | jq

Example output

{
  "object": "list",
  "data": [
    {
      "id": "Llama-2-7b-chat-hf",
      "object": "model",
      "created": 1712919071,
      "owned_by": "fastchat",
      "root": "Llama-2-7b-chat-hf",
      "parent": null,
      "permission": [
        {
          "id": "modelperm-XpFyEE7Sewx4XYbEcdbCVz",
          "object": "model_permission",
          "created": 1712919071,
          "allow_create_engine": false,
          "allow_sampling": true,
          "allow_logprobs": true,
          "allow_search_indices": true,
          "allow_view": true,
          "allow_fine_tuning": false,
          "organization": "*",
          "group": null,
          "is_blocking": false
        }
      ]
    }
  ]
}

Chat Completions

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Llama-2-7b-chat-hf",
    "messages": [{"role": "user", "content": "Hello! What is your name?"}]
  }' | jq

Example output

{
  "id": "chatcmpl-jJ9vKSGkcDMTxKfLxK7q2x",
  "object": "chat.completion",
  "created": 1712919092,
  "model": "Llama-2-7b-chat-hf",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": " Hello! My name is LLaMA, I'm a large language model trained by a team of researcher at Meta AI. Unterscheidung. 😊"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 15,
    "total_tokens": 53,
    "completion_tokens": 38
  }
}

Text Completions

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Llama-2-7b-chat-hf",
    "prompt": "Once upon a time",
    "max_tokens": 41,
    "temperature": 0.5
  }' | jq

Example Output:

{
  "id": "cmpl-PsAkpTWMmBLzWCTtM4r97Y",
  "object": "text_completion",
  "created": 1712919307,
  "model": "Llama-2-7b-chat-hf",
  "choices": [
    {
      "index": 0,
      "text": ", in a far-off land, there was a magical kingdom called \"Happily Ever Laughter.\" It was a place where laughter was the key to happiness, and everyone who ",
      "logprobs": null,
      "finish_reason": "length"
    }
  ],
  "usage": {
    "prompt_tokens": 5,
    "total_tokens": 45,
    "completion_tokens": 40
  }
}