ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[LLM] Use to build a more slim docker for k8s (#9608)

Browse source 
* Create Dockerfile.k8s

* Update Dockerfile

More slim standalone image

* Update Dockerfile

* Update Dockerfile.k8s

* Update bigdl-qlora-finetuing-entrypoint.sh

* Update qlora_finetuning_cpu.py

* Update alpaca_qlora_finetuning_cpu.py

Refer to this [pr](https://github.com/intel-analytics/BigDL/pull/9551/files#diff-2025188afa54672d21236e6955c7c7f7686bec9239532e41c7983858cc9aaa89), update the LoraConfig

* update

* update

* update

* update

* update

* update

* update

* update transformer version

* update Dockerfile

* update Docker image name

* fix error
This commit is contained in:
ZehuaCao 2023-12-08 10:25:36 +08:00 committed by GitHub
parent 6eca8a8bb5
commit d204125e88
5 changed files with 119 additions and 40 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

50
docker/llm/finetune/qlora/cpu/docker/Dockerfile
View file

@ -18,57 +18,39 @@ ENV TRANSFORMERS_COMMIT_ID=95fe0f5
COPY --from=key-getter /root/intel-oneapi-archive-keyring.gpg /usr/share/keyrings/intel-oneapi-archive-keyring.gpg
RUN echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " > /etc/apt/sources.list.d/oneAPI.list
RUN mkdir -p /bigdl/data && mkdir -p /bigdl/model && \
# install pytorch 2.1.0
# install pytorch 2.1.0
apt-get update && \
apt-get install -y python3-pip python3.9-dev python3-wheel git software-properties-common && \
apt-get install -y --no-install-recommends python3-pip python3.9-dev python3-wheel python3.9-distutils git software-properties-common && \
apt-get clean && \
rm -rf /var/lib/apt/lists/* && \
pip3 install --upgrade pip && \
export PIP_DEFAULT_TIMEOUT=100 && \
pip install --upgrade torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu && \
# install CPU bigdl-llm
pip3 install --pre --upgrade bigdl-llm[all] -i https://pypi.tuna.tsinghua.edu.cn/simple/ && \
# install ipex and oneccl
# install CPU bigdl-llm
pip3 install --pre --upgrade bigdl-llm[all] && \
# install ipex and oneccl
pip install intel_extension_for_pytorch==2.0.100 && \
pip install oneccl_bind_pt -f https://developer.intel.com/ipex-whl-stable && \
# install huggingface dependencies
pip install datasets https://files.pythonhosted.org/packages/9a/06/e4ec2a321e57c03b7e9345d709d554a52c33760e5015fdff0919d9459af0/transformers-4.35.0-py3-none-any.whl && \
# install huggingface dependencies
pip install datasets transformers==4.35.0 && \
pip install fire peft==0.5.0 && \
pip install accelerate==0.23.0 && \
# install basic dependencies
apt-get install -y curl wget git gnupg gpg-agent software-properties-common libunwind8-dev vim less && \
# install python 3.9
ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone && \
env DEBIAN_FRONTEND=noninteractive apt-get update && \
add-apt-repository ppa:deadsnakes/ppa -y && \
apt-get install -y python3.9 && \
rm /usr/bin/python3 && \
ln -s /usr/bin/python3.9 /usr/bin/python3 && \
# install basic dependencies
apt-get update && apt-get install -y curl wget gpg gpg-agent software-properties-common libunwind8-dev && \
# get qlora example code
ln -s /usr/bin/python3 /usr/bin/python && \
apt-get install -y python3-pip python3.9-dev python3-wheel python3.9-distutils && \
curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py && \
# install OpenSSH for MPI to communicate between containers
apt-get install -y --no-install-recommends openssh-client openssh-server && \
mkdir -p /var/run/sshd && \
# allow OpenSSH to talk to containers without asking for confirmation
# by disabling StrictHostKeyChecking.
# mpi-operator mounts the .ssh folder from a Secret. For that to work, we need
# to disable UserKnownHostsFile to avoid write permissions.
# disabling StrictModes avoids directory and files read permission checks.
sed -i 's/[ #]\(.*StrictHostKeyChecking \).*/ \1no/g' /etc/ssh/ssh_config && \
echo " UserKnownHostsFile /dev/null" >> /etc/ssh/ssh_config && \
sed -i 's/#\(StrictModes \).*/\1no/g' /etc/ssh/sshd_config && \
# add bigdl-llm qlora cpu example
cd /bigdl && \
git clone https://github.com/intel-analytics/BigDL.git && \
mv BigDL/python/llm/example/CPU/QLoRA-FineTuning/* . && \
rm -r BigDL
rm -r BigDL && \
chown -R mpiuser /bigdl
# for docker directly run example
COPY ./start-qlora-finetuning-on-cpu.sh /bigdl/start-qlora-finetuning-on-cpu.sh
# for k8s
COPY ./bigdl-qlora-finetuing-entrypoint.sh /bigdl/bigdl-qlora-finetuing-entrypoint.sh
RUN chown -R mpiuser /bigdl
USER mpiuser
ENTRYPOINT ["/bin/bash"]

73
docker/llm/finetune/qlora/cpu/docker/Dockerfile.k8s Normal file
View file

@ -0,0 +1,73 @@
# USE TO BUILD A MORE SLIM IMAGE FOR K8S
FROM ubuntu:20.04 as key-getter
ARG http_proxy
ARG https_proxy
RUN apt-get update && \
apt-get install -y curl gpg && \
curl -fsSL https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2023.PUB | gpg --dearmor | tee /root/intel-oneapi-archive-keyring.gpg
FROM debian:bullseye as builder
ARG http_proxy
ARG https_proxy
ENV TZ=Asia/Shanghai
ARG PIP_NO_CACHE_DIR=false
ENV TRANSFORMERS_COMMIT_ID=95fe0f5
# add public key
COPY --from=key-getter /root/intel-oneapi-archive-keyring.gpg /usr/share/keyrings/intel-oneapi-archive-keyring.gpg
RUN echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " > /etc/apt/sources.list.d/oneAPI.list
RUN mkdir -p /bigdl/data && mkdir -p /bigdl/model && \
apt-get update && \
apt install -y --no-install-recommends openssh-server openssh-client libcap2-bin gnupg2 ca-certificates \
python3-pip python3.9-dev python3-wheel python3.9-distutils git software-properties-common && \
apt-get clean && \
rm -rf /var/lib/apt/lists/* && \
mkdir -p /var/run/sshd && \
setcap CAP_NET_BIND_SERVICE=+eip /usr/sbin/sshd && \
apt remove libcap2-bin -y && \
sed -i "s/[ #]\(.*StrictHostKeyChecking \).*/ \1no/g" /etc/ssh/ssh_config && \
echo " UserKnownHostsFile /dev/null" >> /etc/ssh/ssh_config && \
sed -i "s/[ #]\(.*Port \).*/ \1$port/g" /etc/ssh/ssh_config && \
sed -i "s/#\(StrictModes \).*/\1no/g" /etc/ssh/sshd_config && \
sed -i "s/#\(Port \).*/\1$port/g" /etc/ssh/sshd_config && \
useradd -m mpiuser && \
cp -r /etc/ssh/sshd_config /home/mpiuser/.sshd_config && \
echo "Port $port" >> /home/mpiuser/.sshd_config && \
# install pytorch 2.1.0
pip3 install --upgrade pip && \
export PIP_DEFAULT_TIMEOUT=100 && \
pip install --upgrade torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu && \
# install CPU bigdl-llm
pip3 install --pre --upgrade bigdl-llm[all] && \
# install ipex and oneccl
pip install intel_extension_for_pytorch==2.0.100 && \
pip install oneccl_bind_pt -f https://developer.intel.com/ipex-whl-stable && \
# install huggingface dependencies
pip install datasets transformers==4.35.0 && \
pip install fire peft==0.5.0 && \
pip install accelerate==0.23.0 && \
# install basic dependencies
apt-get update && apt-get install -y curl wget gpg gpg-agent software-properties-common libunwind8-dev && \
# Install Intel oneAPI keys.
apt remove -y gnupg2 ca-certificates && \
apt autoremove -y && \
apt update && \
apt install -y --no-install-recommends dnsutils intel-oneapi-mpi && \
rm -rf /var/lib/apt/lists/* && \
# get qlora example code
ln -s /usr/bin/python3 /usr/bin/python && \
cd /bigdl && \
git clone https://github.com/intel-analytics/BigDL.git && \
mv BigDL/python/llm/example/CPU/QLoRA-FineTuning/* . && \
rm -r BigDL && \
chown -R mpiuser /bigdl
# for standalone
COPY ./start-qlora-finetuning-on-cpu.sh /bigdl/start-qlora-finetuning-on-cpu.sh
USER mpiuser
ENTRYPOINT ["/bin/bash"]

33
docker/llm/finetune/qlora/cpu/docker/README.md
View file

@ -7,20 +7,35 @@ The following shows how to fine-tune LLM with Quantization (QLoRA built on BigDL
You can download directly from Dockerhub like:
```bash
docker pull intelanalytics/bigdl-llm-finetune-qlora-cpu:2.4.0-SNAPSHOT
# For standalone
docker pull intelanalytics/bigdl-llm-finetune-qlora-cpu-standalone:2.5.0-SNAPSHOT
# For k8s
docker pull intelanalytics/bigdl-llm-finetune-qlora-cpu-k8s:2.5.0-SNAPSHOT
```
Or build the image from source:
```bash
# For standalone
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-cpu:2.4.0-SNAPSHOT \
-t intelanalytics/bigdl-llm-finetune-qlora-cpu-standalone:2.5.0-SNAPSHOT \
-f ./Dockerfile .
# For k8s
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-cpu-k8s:2.5.0-SNAPSHOT \
-f ./Dockerfile.k8s .
```
### 2. Prepare Base Model, Data and Container
@ -40,7 +55,7 @@ docker run -itd \
-e https_proxy=${HTTPS_PROXY} \
-v $BASE_MODE_PATH:/bigdl/model \
-v $DATA_PATH:/bigdl/data/english_quotes \
intelanalytics/bigdl-llm-finetune-qlora-cpu:2.4.0-SNAPSHOT
intelanalytics/bigdl-llm-finetune-qlora-cpu-standalone:2.5.0-SNAPSHOT
```
The download and mount of base model and data to a docker container demonstrates a standard fine-tuning process. You can skip this step for a quick start, and in this way, the fine-tuning codes will automatically download the needed files:
@ -54,7 +69,7 @@ docker run -itd \
--name=bigdl-llm-fintune-qlora-cpu \
-e http_proxy=${HTTP_PROXY} \
-e https_proxy=${HTTPS_PROXY} \
intelanalytics/bigdl-llm-finetune-qlora-cpu:2.4.0-SNAPSHOT
intelanalytics/bigdl-llm-finetune-qlora-cpu-standalone: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).
@ -93,7 +108,9 @@ TrainOutput(global_step=200, training_loss=1.5072882556915284, metrics={'train_r
```
### 4. Merge the adapter into the original model
Using the [export_merged_model.py](https://github.com/intel-analytics/BigDL/blob/main/python/llm/example/GPU/QLoRA-FineTuning/export_merged_model.py) to merge.
```
python ./export_merged_model.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --adapter_path ./outputs/checkpoint-200 --output_path ./outputs/checkpoint-200-merged
```
@ -101,14 +118,18 @@ python ./export_merged_model.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --
Then you can use `./outputs/checkpoint-200-merged` as a normal huggingface transformer model to do inference.
### 5. Use BigDL-LLM to verify the fine-tuning effect
Train more steps and try input sentence like `['quote'] -> [?]` to verify. For example, using `“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->: ` to inference.
BigDL-LLM llama2 example [link](https://github.com/intel-analytics/BigDL/tree/main/python/llm/example/CPU/HF-Transformers-AutoModels/Model/llama2). Update the `LLAMA2_PROMPT_FORMAT = "{prompt}"`.
```bash
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt "“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->:" --n-predict 20
```
#### Sample Output
Base_model output
```log
Inference time: xxx s
-------------------- Prompt --------------------
@ -116,7 +137,9 @@ Inference time: xxx s
-------------------- Output --------------------
“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->: 💻 Fine-tuning a language model on a powerful device like an Intel CPU
```
Merged_model output
```log
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Inference time: xxx s
@ -140,7 +163,7 @@ docker run -itd \
-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-qlora-cpu:2.5.0-SNAPSHOT
intelanalytics/bigdl-llm-finetune-qlora-cpu-standalone:2.5.0-SNAPSHOT
```
Note that `STANDALONE_DOCKER` is set to **TRUE** here.

1
docker/llm/finetune/qlora/cpu/docker/bigdl-qlora-finetuing-entrypoint.sh
View file

@ -4,6 +4,7 @@ set -x
source /opt/intel/oneapi/setvars.sh
export CCL_WORKER_COUNT=$WORLD_SIZE
source bigdl-llm-init -t
cd /bigdl/alpaca-qlora
if [ "$WORKER_ROLE" = "launcher" ]
then
sed "s/:1/ /g" /etc/mpi/hostfile > /home/mpiuser/hostfile

2
docker/llm/finetune/qlora/cpu/kubernetes/values.yaml
View file

@ -1,4 +1,4 @@
imageName: intelanalytics/bigdl-llm-finetune-qlora-cpu:2.5.0-SNAPSHOT
imageName: intelanalytics/bigdl-llm-finetune-qlora-cpu-k8s:2.5.0-SNAPSHOT
trainerNum: 2
microBatchSize: 8
enableGradientCheckpoint: false # true will save more memory but increase latency