- Docker installation guide for BigDL-LLM on CPU
- Docker installation guide for BigDL LLM on XPU
- Docker installation guide for BigDL LLM Serving on CPU
- Docker installation guide for BigDL LLM Serving on XPU
- Docker installation guide for BigDL LLM Fine Tuning on CPU
- Docker installation guide for BigDL LLM Fine Tuning on XPU
New users can quickly get started with Docker using this official link.
For Windows users, make sure WSL2 or Hyper-V is enabled on your computer. The instructions for installing can be accessed from here.
To pull image from hub, you can execute command on console:
docker pull intelanalytics/bigdl-llm-cpu:2.5.0-SNAPSHOT
to check if the image is successfully downloaded, you can use:
docker images | sls intelanalytics/bigdl-llm-cpu:2.5.0-SNAPSHOT
To run the image and do inference, you could create and run a bat script on Windows.
An example on Windows could be:
@echo off
set DOCKER_IMAGE=intelanalytics/bigdl-llm-cpu:2.5.0-SNAPSHOT
set CONTAINER_NAME=my_container
set MODEL_PATH=D:/llm/models[change to your model path]
:: Run the Docker container
docker run -itd ^
-p 12345:12345 ^
--cpuset-cpus="0-7" ^
--cpuset-mems="0" ^
--memory="8G" ^
--name=%CONTAINER_NAME% ^
-v %MODEL_PATH%:/llm/models ^
%DOCKER_IMAGE%
After the container is booted, you could get into the container through docker exec
.
docker exec -it my_container bash
To run inference using BigDL-LLM
using cpu, you could refer to this documentation.
chat.py can be used to initiate a conversation with a specified model. The file is under directory '/llm'.
You can download models and bind the model directory from host machine to container when start a container.
After entering the container through docker exec
, you can run chat.py by:
cd /llm/portable-zip
python chat.py --model-path YOUR_MODEL_PATH
If your model is chatglm-6b and mounted on /llm/models, you can excute:
python chat.py --model-path /llm/models/chatglm-6b
Here is a demostration:
You could start a jupyter-lab serving to explore bigdl-llm-tutorial which can help you build a more sophisticated Chatbo.
To start serving, run the script under '/llm':
cd /llm
./start-notebook.sh [--port EXPECTED_PORT]
You could assign a port to serving, or the default port 12345 will be assigned.
If you use host network mode when booted the container, after successfully running service, you can access http://127.0.0.1:12345/lab to get into tutorial, or you should bind the correct ports between container and host.
Here is a demostration of how to use tutorial in explorer:
To run container on Linux/MacOS:
#/bin/bash
export DOCKER_IMAGE=intelanalytics/bigdl-llm-cpu:2.5.0-SNAPSHOT
export CONTAINER_NAME=my_container
export MODEL_PATH=/llm/models[change to your model path]
docker run -itd \
-p 12345:12345 \
--cpuset-cpus="0-7" \
--cpuset-mems="0" \
--memory="8G" \
--name=$CONTAINER_NAME \
-v $MODEL_PATH:/llm/models \
$DOCKER_IMAGE
Also, you could use chat.py and bigdl-llm-tutorial for development.
Getting started with tutorials
First, pull docker image from docker hub:
docker pull intelanalytics/bigdl-llm-xpu:2.5.0-SNAPSHOT
To map the xpu into the container, you need to specify --device=/dev/dri when booting the container. An example could be:
#/bin/bash
export DOCKER_IMAGE=intelanalytics/bigdl-llm-xpu:2.5.0-SNAPSHOT
export CONTAINER_NAME=my_container
export MODEL_PATH=/llm/models[change to your model path]
sudo docker run -itd \
--net=host \
--device=/dev/dri \
--memory="32G" \
--name=$CONTAINER_NAME \
--shm-size="16g" \
-v $MODEL_PATH:/llm/models \
$DOCKER_IMAGE
After the container is booted, you could get into the container through docker exec
.
To verify the device is successfully mapped into the container, run sycl-ls
to check the result. In a machine with Arc A770, the sampled output is:
root@arda-arc12:/# sycl-ls
[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device 1.2 [2023.16.7.0.21_160000]
[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i9-13900K 3.0 [2023.16.7.0.21_160000]
[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics 3.0 [23.17.26241.33]
[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26241]
To run inference using BigDL-LLM
using xpu, you could refer to this documentation.
Pull image:
docker pull intelanalytics/bigdl-llm-serving-cpu:2.5.0-SNAPSHOT
You could use the following bash script to start the container. Please be noted that the CPU config is specified for Xeon CPUs, change it accordingly if you are not using a Xeon CPU.
export DOCKER_IMAGE=intelanalytics/bigdl-llm-serving-cpu:2.5.0-SNAPSHOT
export CONTAINER_NAME=my_container
export MODEL_PATH=/llm/models[change to your model path]
docker run -itd \
--net=host \
--cpuset-cpus="0-47" \
--cpuset-mems="0" \
--memory="32G" \
--name=$CONTAINER_NAME \
-v $MODEL_PATH:/llm/models \
$DOCKER_IMAGE
After the container is booted, you could get into the container through docker exec
.
Using BigDL-LLM in FastChat does not impose any new limitations on model usage. Therefore, all Hugging Face Transformer models can be utilized in FastChat.
FastChat determines the Model adapter to use through path matching. Therefore, in order to load models using BigDL-LLM, you need to make some modifications to the model's name.
For instance, assuming you have downloaded the llama-7b-hf
from HuggingFace. Then, to use the BigDL-LLM
as backend, you need to change name from llama-7b-hf
to bigdl-7b
.
The key point here is that the model's path should include "bigdl" and should not include paths matched by other model adapters.
A special case is ChatGLM
models. For these models, you do not need to do any changes after downloading the model and the BigDL-LLM
backend will be used automatically.
To serve using the Web UI, you need three main components: web servers that interface with users, model workers that host one or more models, and a controller to coordinate the web server and model workers.
python3 -m fastchat.serve.controller
This controller manages the distributed workers.
python3 -m bigdl.llm.serving.model_worker --model-path lmsys/vicuna-7b-v1.3 --device cpu
Wait until the process finishes loading the model and you see "Uvicorn running on ...". The model worker will register itself to the controller.
To run model worker using Intel GPU, simply change the --device cpu option to --device xpu
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 BigDL-LLM
as the backend. You can open your browser and chat with a model now.
To start an OpenAI API server that provides compatible APIs using BigDL-LLM
backend, you need three main components: an OpenAI API Server that serves the in-coming requests, model workers that host one or more models, and a controller to coordinate the web server and model workers.
First, launch the controller
python3 -m fastchat.serve.controller
Then, launch the model worker(s):
python3 -m bigdl.llm.serving.model_worker --model-path lmsys/vicuna-7b-v1.3 --device cpu
Finally, launch the RESTful API server
python3 -m fastchat.serve.openai_api_server --host localhost --port 8000
Pull image:
docker pull intelanalytics/bigdl-llm-serving-xpu:2.5.0-SNAPSHOT
To map the xpu
into the container, you need to specify --device=/dev/dri
when booting the container.
An example could be:
#/bin/bash
export DOCKER_IMAGE=intelanalytics/bigdl-llm-serving-cpu:2.5.0-SNAPSHOT
export CONTAINER_NAME=my_container
export MODEL_PATH=/llm/models[change to your model path]
export SERVICE_MODEL_PATH=/llm/models/chatglm2-6b[a specified model path for running service]
docker run -itd \
--net=host \
--device=/dev/dri \
--memory="32G" \
--name=$CONTAINER_NAME \
--shm-size="16g" \
-v $MODEL_PATH:/llm/models \
-e SERVICE_MODEL_PATH=$SERVICE_MODEL_PATH \
$DOCKER_IMAGE --service-model-path $SERVICE_MODEL_PATH
You can assign specified model path to service-model-path to run the service while booting the container. Also you can manually run the service after entering container. Run /opt/entrypoint.sh --help
in container to see more information. There are steps below describe how to run service in details as well.
To verify the device is successfully mapped into the container, run sycl-ls
to check the result. In a machine with Arc A770, the sampled output is:
root@arda-arc12:/# sycl-ls
[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device 1.2 [2023.16.7.0.21_160000]
[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i9-13900K 3.0 [2023.16.7.0.21_160000]
[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics 3.0 [23.17.26241.33]
[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26241]
After the container is booted, you could get into the container through docker exec
.
To serve using the Web UI, you need three main components: web servers that interface with users, model workers that host one or more models, and a controller to coordinate the web server and model workers.
python3 -m fastchat.serve.controller
This controller manages the distributed workers.
python3 -m bigdl.llm.serving.model_worker --model-path lmsys/vicuna-7b-v1.3 --device xpu
Wait until the process finishes loading the model and you see "Uvicorn running on ...". The model worker will register itself to the controller.
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 BigDL-LLM
as the backend. You can open your browser and chat with a model now.
To start an OpenAI API server that provides compatible APIs using BigDL-LLM
backend, you need three main components: an OpenAI API Server that serves the in-coming requests, model workers that host one or more models, and a controller to coordinate the web server and model workers.
First, launch the controller
python3 -m fastchat.serve.controller
Then, launch the model worker(s):
python3 -m bigdl.llm.serving.model_worker --model-path lmsys/vicuna-7b-v1.3 --device xpu
Finally, launch the RESTful API server
python3 -m fastchat.serve.openai_api_server --host localhost --port 8000
You can download directly from Dockerhub like:
docker pull intelanalytics/bigdl-llm-finetune-lora-cpu:2.5.0-SNAPSHOT
Or build the image from source:
export HTTP_PROXY=your_http_proxy
export HTTPS_PROXY=your_https_proxy
docker build \
--build-arg http_proxy=${HTTP_PROXY} \
--build-arg https_proxy=${HTTPS_PROXY} \
-t intelanalytics/bigdl-llm-finetune-lora-cpu:2.5.0-SNAPSHOT \
-f ./Dockerfile .
Here, we try to finetune Llama2-7b with Cleaned alpaca data, which contains all kinds of general knowledge and has already been cleaned. And please download them and start a docker container with files mounted like below:
docker run -itd \
--name=bigdl-llm-fintune-lora-cpu \
--cpuset-cpus="your_expected_range_of_cpu_numbers" \
-e STANDALONE_DOCKER=TRUE \
-e WORKER_COUNT_DOCKER=your_worker_count \
-v your_downloaded_base_model_path:/bigdl/model \
-v your_downloaded_data_path:/bigdl/data/alpaca_data_cleaned_archive.json \
intelanalytics/bigdl-llm-finetune-cpu:2.5.0-SNAPSHOT \
bash
You can adjust the configuration according to your own environment. After our testing, we recommend you set worker_count=1, and then allocate 80G memory to Docker.
Enter the running container:
docker exec -it bigdl-llm-fintune-lora-cpu bash
Then, run the script to start finetuning:
bash /bigdl/bigdl-lora-finetuing-entrypoint.sh
After minutes, it is expected to get results like:
Training Alpaca-LoRA model with params:
...
Related params
...
world_size: 2!!
PMI_RANK(local_rank): 1
Loading checkpoint shards: 100%|██████████| 2/2 [00:04<00:00, 2.28s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:05<00:00, 2.62s/it]
trainable params: 4194304 || all params: 6742609920 || trainable%: 0.06220594176090199
[INFO] spliting and shuffling dataset...
[INFO] shuffling and tokenizing train data...
Map: 2%|▏ | 1095/49759 [00:00<00:30, 1599.00 examples/s]trainable params: 4194304 || all params: 6742609920 || trainable%: 0.06220594176090199
[INFO] spliting and shuffling dataset...
[INFO] shuffling and tokenizing train data...
Map: 100%|██████████| 49759/49759 [00:29<00:00, 1678.89 examples/s]
[INFO] shuffling and tokenizing test data...
Map: 100%|██████████| 49759/49759 [00:29<00:00, 1685.42 examples/s]
[INFO] shuffling and tokenizing test data...
Map: 100%|██████████| 2000/2000 [00:01<00:00, 1573.61 examples/s]
Map: 100%|██████████| 2000/2000 [00:01<00:00, 1578.71 examples/s]
[INFO] begining the training of transformers...
[INFO] Process rank: 0, device: cpudistributed training: True
0%| | 1/1164 [02:42<52:28:24, 162.43s/it]
You can run BF16-Optimized lora finetuning on kubernetes with OneCCL. So for kubernetes users, please refer to here.
The following shows how to fine-tune LLM with Quantization (QLoRA built on BigDL-LLM 4bit optimizations) in a docker environment, which is accelerated by Intel XPU.
You can download directly from Dockerhub like:
docker pull intelanalytics/bigdl-llm-finetune-qlora-xpu:2.5.0-SNAPSHOT
Or build the image from source:
export HTTP_PROXY=your_http_proxy
export HTTPS_PROXY=your_https_proxy
docker build \
--build-arg http_proxy=${HTTP_PROXY} \
--build-arg https_proxy=${HTTPS_PROXY} \
-t intelanalytics/bigdl-llm-finetune-qlora-xpu:2.5.0-SNAPSHOT \
-f ./Dockerfile .
Here, we try to fine-tune a Llama2-7b with yahma/alpaca-cleaned dataset, and please download them and start a docker container with files mounted like below:
export BASE_MODE_PATH=your_downloaded_base_model_path
export DATA_PATH=your_downloaded_data_path
export HTTP_PROXY=your_http_proxy
export HTTPS_PROXY=your_https_proxy
docker run -itd \
--net=host \
--device=/dev/dri \
--memory="32G" \
--name=bigdl-llm-fintune-qlora-xpu \
-e http_proxy=${HTTP_PROXY} \
-e https_proxy=${HTTPS_PROXY} \
-v $BASE_MODE_PATH:/model \
-v $DATA_PATH:/data/alpaca-cleaned \
--shm-size="16g" \
intelanalytics/bigdl-llm-fintune-qlora-xpu:2.5.0-SNAPSHOT
The download and mount of base model and data to a docker container demonstrates a standard fine-tuning process. You can skip this step for a quick start, and in this way, the fine-tuning codes will automatically download the needed files:
export HTTP_PROXY=your_http_proxy
export HTTPS_PROXY=your_https_proxy
docker run -itd \
--net=host \
--device=/dev/dri \
--memory="32G" \
--name=bigdl-llm-fintune-qlora-xpu \
-e http_proxy=${HTTP_PROXY} \
-e https_proxy=${HTTPS_PROXY} \
--shm-size="16g" \
intelanalytics/bigdl-llm-fintune-qlora-xpu:2.5.0-SNAPSHOT
However, we do recommend you to handle them manually, because the automatical download can be blocked by Internet access and Huggingface authentication etc. according to different environment, and the manual method allows you to fine-tune in a custom way (with different base model and dataset).
Enter the running container:
docker exec -it bigdl-llm-fintune-qlora-xpu bash
Then, start QLoRA fine-tuning:
bash start-qlora-finetuning-on-xpu.sh
After minutes, it is expected to get results like:
{'loss': 2.0251, 'learning_rate': 0.0002, 'epoch': 0.02}
{'loss': 1.2389, 'learning_rate': 0.00017777777777777779, 'epoch': 0.03}
{'loss': 1.032, 'learning_rate': 0.00015555555555555556, 'epoch': 0.05}
{'loss': 0.9141, 'learning_rate': 0.00013333333333333334, 'epoch': 0.06}
{'loss': 0.8505, 'learning_rate': 0.00011111111111111112, 'epoch': 0.08}
{'loss': 0.8713, 'learning_rate': 8.888888888888889e-05, 'epoch': 0.09}
{'loss': 0.8635, 'learning_rate': 6.666666666666667e-05, 'epoch': 0.11}
{'loss': 0.8853, 'learning_rate': 4.4444444444444447e-05, 'epoch': 0.12}
{'loss': 0.859, 'learning_rate': 2.2222222222222223e-05, 'epoch': 0.14}
{'loss': 0.8608, 'learning_rate': 0.0, 'epoch': 0.15}
{'train_runtime': xxxx, 'train_samples_per_second': xxxx, 'train_steps_per_second': xxxx, 'train_loss': 1.0400420665740966, 'epoch': 0.15}
100%|███████████████████████████████████████████████████████████████████████████████████| 200/200 [07:16<00:00, 2.18s/it]
TrainOutput(global_step=200, training_loss=1.0400420665740966, metrics={'train_runtime': xxxx, 'train_samples_per_second': xxxx, 'train_steps_per_second': xxxx, 'train_loss': 1.0400420665740966, 'epoch': 0.15})