Skip to content

README.md

Latest commit

 

History

History
177 lines (129 loc) · 4.67 KB

README.md

File metadata and controls

177 lines (129 loc) · 4.67 KB

FrontierAI

AI related notes on Frontier supercomputer at OLCF

Contents

FlashAttention

Installation

FA2 is supported on Frontier and the upstream repo can be pip installed

module load PrgEnv-gnu
module load rocm
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O Miniconda3-latest-Linux-x86_64.sh
bash ./Miniconda3-latest-Linux-x86_64.sh -b -p $WRKSPC/miniconda
export PATH=$WRKSPC/miniconda/bin:$PATH
conda create --prefix $WRKSPC/miniconda/envs/fa2-env -y
source $WRKSPC/miniconda/etc/profile.d/conda.sh
conda activate $WRKSPC/miniconda/envs/fa2-env
git clone https://github.com/Dao-AILab/flash-attention
pushd flash-attention
git checkout v2.8.3
pip install -e .
popd

For latest development, please try AMD's fork

Backend

Standalone FA:

  • CK (default)
  • Triton

PyTorch scaled_dot_product_attention (SDPA):

  • Math
  • FA
  • Efficient

Performance

  • For standalone FA, use latest rocm. The built against rocm/6.3 is 1.5x faster than that with rocm/6.1 for certain inputs FA2

  • For PyTorch SDPA, use FA or Efficient backend
    SDPA

X-MoE

Official X-MoE Code and Documentation

https://github.com/Supercomputing-System-AI-Lab/X-MoE

Preparing Conda Environment

module reset
module load cpe/24.11
module load PrgEnv-gnu/8.6.0
module load rocm/6.3.1
module load craype-accel-amd-gfx90a
module load miniforge3/23.11.0-0


# cd to your directory of choice,
# I recommend using /lustre/orion directories as these packages require lots of space.

conda create -p $PWD/TORCH-ROCM6.3.1_env python=3.11 -c conda-forge -y
source activate $PWD/TORCH-ROCM6.3.1_env

Installing Torch

# ROCM 6.3.1
pip3 install --pre torch torchvision torchaudio --index-url
https://download.pytorch.org/whl/nightly/rocm6.3

Installing APEX

git clone https://github.com/ROCm/apex.git
cd apex
pip install -r requirements.txt
python setup.py install --cpp_ext --cuda_ext

Installing FlashAttention

Didn't work with pip install, so tried to install from source instead.

git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention/
python setup.py install
pytest -q -s tests/test_flash_attn.py

Installing X-MoE

Worked as is.

cd ~
git clone https://github.com/Supercomputing-System-AI-Lab/X-MoE
cd X-MoE
git submodule update --init --recursive --remote

pip install -e .
cd Megatron-DeepSpeed-X-MoE && pip install -e .

Data Preparation

Recommending removing   from XMoE's parent's directory. Worked as is.

Maybe make the amount of data a variable so that we don't have to go through the entire dataset.

Training on Single Node

Didn't work on login node. Looks like this depends on having a host list.

scontrol: error: host list is empty
first=
ssh: Could not resolve hostname : Name or service not known
MASTER_ADDR=

Getting an Interactive Node

    salloc -A PROJID -J RunSim123 -t 0:30:00 -p batch -N 1

Running the Training

    ./X-MoE-Small-node-1.sh 8 1

Frontier pytorch $cpp_extension$ related problem, "ninja -v" needs to be replaced with

    ninja --version

.

Need to install mpi4py

    MPICC="cc -shared" pip install --no-cache-dir --no-binary=mpi4py mpi4py

Replaced torchrun with srun

    time srun -u -N 1 -n${NUM_GPUS} -c2 --ntasks-per-node=8
    --gpus-per-node=8 --gpu-bind=closest python pretrain_gpt_deepspeed.py ...

Error Logs

./X-MoE-Small-node-1-srun.sh 8 1

srun_logs_with_jit_error.log

     ImportError: /lustre/orion/stf218/world-shared/sajal/testing_xmoe/
     X-MoE/Megatron-DeepSpeed-X-MoE/
     megatron/fused_kernels/build/scaled_upper_triang_masked_softmax_cuda.so:
     cannot open shared object file: No such file or directory

Using Correct Versions of Compiler

    export CXX=/opt/cray/pe/gcc-native/13/bin/g++
    export CC=/opt/cray/pe/gcc-native/13/bin/gcc
    export PATH=/opt/cray/pe/gcc-native/13/bin:$PATH