llm_finetune

Running Intel-HabanaLabs MLPerf™ Llama-70B LoRA Benchmark

This directory provides instructions to reproduce Intel-HabanaLabs's results for MLPerf Training v4.0 Llama-70B LoRA benchmark on single server with 8 Gaudi2 cards.

For more information on training deep learning models using Gaudi, refer to developer.habana.ai

MLPerf™ is a trademark and service mark of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited.

• Running Intel-HabanaLabs MLPerf™ Llama-70B LoRA Benchmark
  • Setup
    • Prepare MLPerf Directory
    • Build and Deploy Intel-HabanaLabs MLPerf Training 4.0 Container
    • Download Data and Model
  • Finetuning Llama2 70B with LoRA

Setup

Make sure to have requested permission for donwloading Llama2 weights on the Hugging Face Hub: https://huggingface.co/meta-llama/Llama-2-7b-hf

Prepare MLPerf Directory

On each compute node, perform the following:

1. Follow the instructions provided in the Gaudi Installation Guide to set up the environment including the $PYTHON environment variable. The guide will walk you through the process of setting up your system to run the benchmarks on Gaudi.

2. Create directories for dataset and logs:

   export MLPERF_DIR=/path/to/mlperf/root
   export DATASETS_DIR=/path/to/datasets
   export MODEL_DIR=/path/to/model
   mkdir -p $MLPERF_DIR/Intel-HabanaLabs $MODEL_DIR $DATASETS_DIR
   

3. This README is located in benchmarks/llm_finetune directory corresponding to Intel-HabanaLabs's Llama-70B LoRA submission. Download this whole benchmarks folder along with all subfolders and copy it under $MLPERF_DIR/Intel-HabanaLabs

Build and Deploy Intel-HabanaLabs MLPerf Training 4.0 Container

1. Create mlperf4.0 container by running the following command.

• TODO: update DOCKER_IMAGE once it is known and published.

  export CONTAINER_NAME=mlperf4.0
  export DOCKER_IMAGE=vault.habana.ai/gaudi-docker-mlperf/ver4.0/pytorch-installer-2.2.0:1.16.98-46
  docker run --privileged --security-opt seccomp=unconfined \
    --name $CONTAINER_NAME -td                              \
    -v /dev:/dev                                            \
    --device=/dev:/dev                                      \
    -e LOG_LEVEL_ALL=6                                      \
    -v /sys/kernel/debug:/sys/kernel/debug                  \
    -v /tmp:/tmp                                            \
    -v $MLPERF_DIR:/root/MLPERF                             \
    -v $DATASETS_DIR:/root/datasets                         \
    -v $MODEL_DIR:/root/model                               \
    --cap-add=sys_nice --cap-add=SYS_PTRACE                 \
    --user root --workdir=/root --net=host                  \
    --ulimit memlock=-1:-1 ${DOCKER_IMAGE}

2. Start the docker.

   docker exec $CONTAINER_NAME bash -c "service ssh start"
   docker exec -it $CONTAINER_NAME bash

Download Data and Model

MLCommons hosts the model for download exclusively by MLCommons Members. You must first agree to the confidentiality notice, then follow the [link[(https://drive.google.com/drive/folders/11tBZvvrh0FCm3XuR5E849K42TqftYdUF)] to a directory containing Rclone download instructions. Follow steps 1-3 to install and activate Rclone. Finally, download the model to the desired download directory (default ./models): Log into mlperf4.0 container and run:

rclone copy mlc-llama2:Llama2-70b-fused-qkv-mlperf /root/model/Llama2-70b-fused-qkv-mlperf -P

Similarly download the data to the desired download directory (default ./dataset):

rclone copy mlc-llama2:training/scrolls_gov_report_8k /root/datasets/scrolls_gov_report_8k -P

Finetuning Llama2 70B with LoRA

1. Inside the mlperf4.0 container, install requirements:

pip install git+https://github.com/HabanaAI/[email protected]
pip install git+https://github.com/HabanaAI/optimum-habana-fork.git@cef6209
pip install -r  /root/MLPERF/Intel-HabanaLabs/benchmarks/llm_finetune/requirements.txt
huggingface-cli login

2. Create device warmup data:

cd /root/datasets/scrolls_gov_report_8k
python /root/MLPERF/Intel-HabanaLabs/benchmarks/llm_finetune/scripts/create_warmup_data.py

3. Run the training.

cd /root/MLPERF/Intel-HabanaLabs/benchmarks/llm_finetune/
cp /root/MLPERF/Intel-HabanaLabs/benchmarks/llm_finetune/config.json /root/model/Llama2-70b-fused-qkv-mlperf/
./run_llama_70B_fp8_submission.sh