it4i logoDocumentation
Lumi/Lumi soft

PyTorch

PyTorch Highlight

CSC Installed Software Collection

PyTorch Install

Base Environment

module purge
module load CrayEnv
module load PrgEnv-cray/8.3.3
module load craype-accel-amd-gfx90a
module load cray-python

# Default ROCm – more recent versions are preferable (e.g. ROCm 5.6.0)
module load rocm/5.2.3.lua

Scripts

PyTorch Tests

Run Interactive Job on Single Node

salloc -A project_XXX --partition=standard-g -N 1 -n 1 --gpus 8 -t 01:00:00

Scripts

Run Interactive Job on Multiple Nodes

salloc -A project_XXX --partition=standard-g -N 2 -n 16 --gpus 16 -t 01:00:00

Scripts

Tips

Official Containers

ls -la /appl/local/containers/easybuild-sif-images/

Unofficial Versions of ROCM

module use /pfs/lustrep2/projappl/project_462000125/samantao-public/mymodules
ml rocm/5.4.3
ml rocm/5.6.0

Unofficial Containers

ls -la /pfs/lustrep2/projappl/project_462000125/samantao-public/containers/

Installing Python Modules in Containers

#!/bin/bash

wd=$(pwd)
SIF=/pfs/lustrep2/projappl/project_462000125/samantao-public/containers/lumi-pytorch-rocm-5.6.1-python-3.10-pytorch-v2.1.0-dockerhash-aa8dbea5e0e4.sif

rm -rf $wd/setup-me.sh
cat > $wd/setup-me.sh << EOF
#!/bin/bash -e

\$WITH_CONDA
pip3 install scipy h5py tqdm
EOF
chmod +x $wd/setup-me.sh

mkdir -p $wd/pip_install

srun -n 1 --gpus 8 singularity exec \
-B /var/spool/slurmd:/var/spool/slurmd \
-B /opt/cray:/opt/cray \
-B /usr/lib64/libcxi.so.1:/usr/lib64/libcxi.so.1 \
-B $wd:/workdir \
-B $wd/pip_install:$HOME/.local/lib \
$SIF /workdir/setup-me.sh

# Add the path of pip_install to singularity-exec in run.sh:
# -B $wd/pip_install:$HOME/.local/lib \

Controlling Device Visibility

  • HIP_VISIBLE_DEVICES=0,1,2,3 python -c 'import torch; print(torch.cuda.device_count())'
  • ROCR_VISIBLE_DEVICES=0,1,2,3 python -c 'import torch; print(torch.cuda.device_count())'
  • SLURM sets ROCR_VISIBLE_DEVICES
  • Implications of both ways of setting visibility – blit kernels and/or DMA

RCCL

  • The problem – on startup we can see:
    • NCCL error in: /pfs/lustrep2/projappl/project_462000125/samantao/pytorchexample/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1269, unhandled system error, NCCL version 2.12.12
  • Checking error origin:
    • export NCCL_DEBUG=INFO
    • NCCL INFO NET/Socket : Using [0]nmn0:10.120.116.65<0> [1]hsn0:10.253.6.67<0> [2]hsn1:10.253.6.68<0>[3]hsn2:10.253.2.12<0> [4]hsn3:10.253.2.11<0>
    • NCCL INFO /long_pathname_so_that_rpms_can_package_the_debug_info/data/driver/rccl/src/init.cc:1292
  • The fix:
    • export NCCL_SOCKET_IFNAME=hsn0,hsn1,hsn2,hsn3

RCCL AWS-CXI Plugin

  • RCCL relies on runtime plugin-ins to connect with some transport layers
    • Libfabric – provider for Slingshot
  • Hipified plugin adapted from AWS OpenFabrics support available
  • https://github.com/ROCmSoftwarePlatform/aws-ofi-rccl
  • 3-4x faster collectives
  • Plugin needs to be pointed at by the loading environment
module use /pfs/lustrep2/projappl/project_462000125/samantao-public/mymodules
module load aws-ofi-rccl/rocm-5.2.3.lua
# Or
export LD_LIBRARY_PATH=/pfs/lustrep2/projappl/project_462000125/samantao-public/apps-rocm-5.2.3/aws-ofirccl
# (will detect librccl-net.so)
  • Verify the plugin is detected
export NCCL_DEBUG=INFO
export NCCL_DEBUG_SUBSYS=INIT
# and search the logs for:
# [0] NCCL INFO NET/OFI Using aws-ofi-rccl 1.4.0

amdgpu.ids Issue

https://github.com/pytorch/builder/issues/1410

References

© 2025 IT4Innovations – All rights reserved.