Skip to content

intel/torch-ccl

Repository files navigation

Intel® oneCCL Bindings for PyTorch (formerly known as torch_ccl)

This repository holds PyTorch bindings maintained by Intel® for the Intel® oneAPI Collective Communications Library (oneCCL).

Introduction

PyTorch is an open-source machine learning framework.

Intel® oneCCL (collective communications library) is a library for efficient distributed deep learning training, implementing collectives like allreduce, allgather, alltoall. For more information on oneCCL, please refer to the oneCCL documentation.

oneccl_bindings_for_pytorch module implements PyTorch C10D ProcessGroup API and can be dynamically loaded as external ProcessGroup and only works on Linux platform now.

Capability

The table below shows which functions are available for use with CPU / Intel dGPU tensors.

CPU GPU
send ×
recv ×
broadcast
all_reduce
reduce
all_gather
gather
scatter
reduce_scatter
all_to_all
barrier

PyTorch API Align

We recommend using Anaconda as Python package management system. The followings are the corresponding branches (tags) of oneccl_bindings_for_pytorch and supported PyTorch.

torch oneccl_bindings_for_pytorch
master master
v2.3.1 ccl_torch2.3.100
v2.1.0 ccl_torch2.1.400
v2.1.0 ccl_torch2.1.300
v2.1.0 ccl_torch2.1.200
v2.1.0 ccl_torch2.1.100
v2.0.1 ccl_torch2.0.100
v1.13 ccl_torch1.13
v1.12.1 ccl_torch1.12.100
v1.12.0 ccl_torch1.12
v1.11.0 ccl_torch1.11
v1.10.0 ccl_torch1.10
v1.9.0 ccl_torch1.9
v1.8.1 ccl_torch1.8
v1.7.1 ccl_torch1.7
v1.6.0 ccl_torch1.6
v1.5-rc3 beta09

The usage details can be found in the README of corresponding branch.

Requirements

  • Python 3.8 or later and a C++17 compiler

  • PyTorch v2.3.1

Build Option List

The following build options are supported in Intel® oneCCL Bindings for PyTorch*.

Build Option Default Value Description
COMPUTE_BACKEND N/A Set oneCCL COMPUTE_BACKEND, set to dpcpp and use DPC++ compiler to enable support for Intel XPU
USE_SYSTEM_ONECCL OFF Use oneCCL library in system
CCL_PACKAGE_NAME oneccl-bind-pt Set wheel name
ONECCL_BINDINGS_FOR_PYTORCH_BACKEND cpu Set backend
CCL_SHA_VERSION False Add git head sha version into wheel name

Launch Option List

The following launch options are supported in Intel® oneCCL Bindings for PyTorch*.

Launch Option Default Value Description
ONECCL_BINDINGS_FOR_PYTORCH_ENV_VERBOSE 0 Set verbose level in oneccl_bindings_for_pytorch
ONECCL_BINDINGS_FOR_PYTORCH_ENV_WAIT_GDB 0 Set 1 to force the oneccl_bindings_for_pytorch wait for GDB attaching
TORCH_LLM_ALLREDUCE 0 Set 1 to enable this prototype feature for better scale-up performance. This is a prototype feature to provide better scale-up performance by enabling optimized collective algorithms in oneCCL and asynchronous execution in torch-ccl. This feature requires XeLink enabled for cross-cards communication.
CCL_BLOCKING_WAIT 0 Set 1 to enable this prototype feature, which is to control whether collectives execution on XPU is host blocking or non-blocking.
CCL_SAME_STREAM 0 Set 1 to enable this prototype feature, which is to allow using a computation stream as communication stream to minimize overhead for streams synchronization.

Installation

Install from Source

  1. clone the oneccl_bindings_for_pytorch.

    git clone https://github.com/intel/torch-ccl.git && cd torch-ccl
    git submodule sync
    git submodule update --init --recursive
  2. Install oneccl_bindings_for_pytorch

    # for CPU Backend Only
    python setup.py install
    # for XPU Backend: use DPC++ Compiler to enable support for Intel XPU
    # build with oneCCL from third party
    COMPUTE_BACKEND=dpcpp python setup.py install
    # build with oneCCL from basekit
    export INTELONEAPIROOT=${HOME}/intel/oneapi
    USE_SYSTEM_ONECCL=ON COMPUTE_BACKEND=dpcpp python setup.py install

Install Prebuilt Wheel

Wheel files are available for the following Python versions. Please always use the latest release to get started.

Extension Version Python 3.6 Python 3.7 Python 3.8 Python 3.9 Python 3.10 Python 3.11
2.3.100
2.1.400
2.1.300
2.1.200
2.1.100
2.0.100
1.13
1.12.100
1.12.0
1.11.0
1.10.0
python -m pip install oneccl_bind_pt==2.3.100 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/

Note: Please set proxy or update URL address to https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/ if you meet connection issue.

Runtime Dynamic Linking

  • If oneccl_bindings_for_pytorch is built without oneCCL and use oneCCL in system, dynamic link oneCCl from oneAPI basekit (recommended usage):
source $basekit_root/ccl/latest/env/vars.sh

Note: Make sure you have installed basekit when using Intel® oneCCL Bindings for Pytorch* on Intel® GPUs.

  • If oneccl_bindings_for_pytorch is built with oneCCL from third party or installed from prebuilt wheel: Dynamic link oneCCL and Intel MPI libraries:
source $(python -c "import oneccl_bindings_for_pytorch as torch_ccl;print(torch_ccl.cwd)")/env/setvars.sh

Dynamic link oneCCL only (not including Intel MPI):

source $(python -c "import oneccl_bindings_for_pytorch as torch_ccl;print(torch_ccl.cwd)")/env/vars.sh

Usage

Note: Please import torch and import intel_extension_for_pytorch, prior to import oneccl_bindings_for_pytorch.

example.py

import torch
import intel_extension_for_pytorch
import oneccl_bindings_for_pytorch
import torch.nn.parallel
import torch.distributed as dist

...

os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = '29500'
os.environ['RANK'] = str(os.environ.get('PMI_RANK', 0))
os.environ['WORLD_SIZE'] = str(os.environ.get('PMI_SIZE', 1))

backend = 'ccl'
dist.init_process_group(backend, ...)
my_rank = dist.get_rank()
my_size = dist.get_world_size()
print("my rank = %d  my size = %d" % (my_rank, my_size))

...

model = torch.nn.parallel.DistributedDataParallel(model, ...)

...

(oneccl_bindings_for_pytorch is built without oneCCL, use oneCCL and MPI(if needed) in system)

source $basekit_root/ccl/latest/env/vars.sh
source $basekit_root/mpi/latest/env/vars.sh

mpirun -n <N> -ppn <PPN> -f <hostfile> python example.py

Performance Debugging

For debugging performance of communication primitives PyTorch's Autograd profiler can be used to inspect time spent inside oneCCL calls.

Example:

profiling.py

import torch.nn.parallel
import torch.distributed as dist
import oneccl_bindings_for_pytorch
import os

os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = '29500'
os.environ['RANK'] = str(os.environ.get('PMI_RANK', 0))
os.environ['WORLD_SIZE'] = str(os.environ.get('PMI_SIZE', 1))

backend = 'ccl'
dist.init_process_group(backend)
my_rank = dist.get_rank()
my_size = dist.get_world_size()
print("my rank = %d  my size = %d" % (my_rank, my_size))

x = torch.ones([2, 2])
y = torch.ones([4, 4])
with torch.autograd.profiler.profile(record_shapes=True) as prof:
    for _ in range(10):
        dist.all_reduce(x)
        dist.all_reduce(y)
dist.barrier()
print(prof.key_averages(group_by_input_shape=True).table(sort_by="self_cpu_time_total"))
mpirun -n 2 -l python profiling.py
[0] my rank = 0  my size = 2
[0] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[0]                                                  Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls          Input Shapes
[0] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[0]                oneccl_bindings_for_pytorch::allreduce        91.41%     297.900ms        91.41%     297.900ms      29.790ms            10              [[2, 2]]
[0]     oneccl_bindings_for_pytorch::wait::cpu::allreduce         8.24%      26.845ms         8.24%      26.845ms       2.684ms            10      [[2, 2], [2, 2]]
[0]     oneccl_bindings_for_pytorch::wait::cpu::allreduce         0.30%     973.651us         0.30%     973.651us      97.365us            10      [[4, 4], [4, 4]]
[0]                oneccl_bindings_for_pytorch::allreduce         0.06%     190.254us         0.06%     190.254us      19.025us            10              [[4, 4]]
[0] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[0] Self CPU time total: 325.909ms
[0]
[1] my rank = 1  my size = 2
[1] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[1]                                                  Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls          Input Shapes
[1] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[1]                oneccl_bindings_for_pytorch::allreduce        96.03%     318.551ms        96.03%     318.551ms      31.855ms            10              [[2, 2]]
[1]     oneccl_bindings_for_pytorch::wait::cpu::allreduce         3.62%      12.019ms         3.62%      12.019ms       1.202ms            10      [[2, 2], [2, 2]]
[1]                oneccl_bindings_for_pytorch::allreduce         0.33%       1.082ms         0.33%       1.082ms     108.157us            10              [[4, 4]]
[1]     oneccl_bindings_for_pytorch::wait::cpu::allreduce         0.02%      56.505us         0.02%      56.505us       5.651us            10      [[4, 4], [4, 4]]
[1] -----------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  --------------------
[1] Self CPU time total: 331.708ms
[1]

Known Issues

For Point-to-point communication, directly call dist.send/recv after initializing the process group in launch script will trigger runtime error. Because all ranks of the group are expected to participate in this call to create communicators in our current implementation, while dist.send/recv only has a pair of ranks' participation. As a result, dist.send/recv should be used after collective call, which ensures all ranks' participation. The further solution for supporting directly call dist.send/recv after initializing the process group is still under investigation.

License

BSD License