Installation:

Build apex==0.1 wheel

1. Clone apex repo

git clone https://github.com/ericharper/apex.git
cd apex
git checkout nm_v1.14.0

2. Replace the contents of the setup.py with the following contents:

import sys
import warnings
import os
from packaging.version import parse, Version

from setuptools import setup, find_packages
import subprocess

import torch
from torch.utils.cpp_extension import BuildExtension, CppExtension, CUDAExtension, CUDA_HOME, load

setup(
    name="apex",
    version="0.1",
    packages=find_packages(
        exclude=("build", "csrc", "include", "tests", "dist", "docs", "tests", "examples", "apex.egg-info",)
    ),
    install_requires=["packaging>20.6",],
    description="PyTorch Extensions written by NVIDIA",
)

3. Build the wheel using the command:

python setup.py bdist_wheel

4. After this, you should see the wheel at dist/. You can use this for installation in next section.

Install the neuron deps:

Install the neuron packages using the command:

pip install --upgrade neuronx-cc==2.* torch-neuronx torchvision --extra-index-url https://pip.repos.neuron.amazonaws.com

Install requirements and neuronx_distributed packages

pip install -r requirements.txt apex/dist/apex-0.1-py3-none-any.whl
pip install neuronx_distributed --extra-index-url https://pip.repos.neuron.amazonaws.com

Build and Install NeuronxDistributedTraining (NxDT)

Let's build the NxDT wheel using the following commands:

./build.sh

This should produce a wheel inside the build directory. You can install that wheel using the command:

pip install build/neuronx_distributed_training-0.1-py3-none-any.whl

How to run.

Setup dataset:

Run the following steps to download the dataset:

wget wget https://raw.githubusercontent.com/aws-neuron/neuronx-distributed/master/examples/training/llama/tp_zero1_llama_hf_pretrain/8B_config_llama3/config.json ~/
wget https://raw.githubusercontent.com/aws-neuron/neuronx-distributed/master/examples/training/llama/get_dataset.py

Download/Tokenize the dataset:

Run the following command to tokenize the dateset:

python get_dataset.py --llama-version 3

If you are working with llama2, then change the version to 2. Note: for the above command to work, you need to download the tokenizer using the following snippet:

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('meta-llama/Meta-Llama-3-8B', token='your_own_hugging_face_token')
# For llama2 uncomment line below
# tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', token='your_own_hugging_face_token')

tokenizer.save_pretrained(".")

After running the get_dataset.py, you should see the dataset been downloaded and tokenized at ~/example_datasets/

Run the example

Clone the examples folder to trn1 instance. Before running training, ensure the config path and dataset path are correctly set inside conf/hf_llama_7B_config.yaml.

Once configured, you can then run the parallel_compile using:

COMPILE=1 CONF_FILE=hf_llama3_8B_config ./train.sh

This should extract all the graphs and compile them in parallel. Once done, you can then run the training job using:

CONF_FILE=hf_llama3_8B_config ./train.sh

Contributing

Formatting code

To format the code, use the following command:

pre-commit run --all-files