vfdev-5

Reproduce issue: pytorch/pytorch#54415

• Download data and extract the data: from https://easyupload.io/4ut2q7

git clone [email protected]:Algomorph/NeuralTracking.git
cd NeuralTracking/
git checkout 6b8a10b2536eda26f9b613c3521b2fe4894602ad

nano .git/config

vfdev-5

# Run it
# torchrun --nproc_per_node=4 example_1.py

import os
import time
import torch
import torch.distributed as dist


def pprint(rank, msg):

vfdev-5

How to setup interactive C++ interpreter with PyTorch C++ library in Jupyter using xeus-cling

Requirements

• installed Jupyter notebook
• installed conda, e.g. /opt/conda
• built PyTorch
  • from source
  • libtorch
  • installed via conda (but can have an issue with _GLIBCXX_USE_CXX11_ABI)

vfdev-5

'''
To run the CPU benchmark: `CUDA_VISIBLE_DEVICES="" python benchmark.py --name cpu`
To run the GPU benchmark: `CUDA_VISIBLE_DEVICES=0 python benchmark.py --name cuda`
To run the distributed benchmark: `python -u -m torch.distributed.launch --nproc_per_node=2 --use_env  benchmark.py --name dist`
'''


import argparse
import time
import math

vfdev-5

How to run tests of PyTorch

Distributed module

• Run all distributed tests

python test/run_test.py -pt -vi distributed/test_distributed

• Run a single test

vfdev-5

PyTorch Tech Share (July 06 2020) - Simple PyTorch distributed computation functionality testing with pytest-xdist.

It is about one of many other approaches on how we can test a custom distributed computation functionality by emulating multiple processes.

What is "distributed setting" in PyTorch ?

• Communications between N application's processes
  • send/receive tensors

vfdev-5

My two cents on debugging, if you have ssh access to some of your nodes, especially the one where an xp failed. There is also a possibility to remotely debug the code. This may seem a bit tricky however could help to check the pipeline. I tested it with VSCode.

• ssh to the machine and the environment (docker container etc)

Using debugpy

1. Install debugpy

pip install debugpy

vfdev-5

FROM python:3.6-buster

# - Install gsutil to copy wheels from gcr
# - Install openblas
# - Download torch & xla
# - Setup Python 3.6 env for Torch XLA wheels
# - Install torch & xla
RUN echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] https://packages.cloud.google.com/apt cloud-sdk main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
  apt-get install -y apt-transport-https ca-certificates gnupg curl && \
  curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \

vfdev-5

#!/usr/bin/env python
# coding: utf-8

import torch
print(torch.__version__)

import ignite
print(ignite.__file__)
print(ignite.__version__)

vfdev-5

#!/bin/bash

# Tests configuration:
if [ -z $version ]; then
    export version="v0.2.1"
#    export version="master"
#    export version="engine_refactor"
    echo "Setup version: $version"
fi