Python bindings for the Alveo Collectives Communication Library (ACCL)
Project description
ACCL Python Bindings
This package provides Python bindings for the Alveo Collective Communication Library (ACCL).
Installation
PyACCL is built and uploaded to PyPI with:
python3 setup.py sdist
python3 -m twine upload dist/*
Install it as follows:
From PyPI
python3 -m pip install pyaccl
Locally from source distribution
python3 -m pip install dist/*.tar.gz
Locally editable
python3 -m pip install -e .
Downloading Notebooks and Overlays
PyACCL provides a few Jupyter notebooks and associated FPGA overlays for Alveo boards. After installing the PyACCL package, you can download these with:
pynq get-notebooks --from-package pyaccl all
Pynq will automatically download the right overlay for the Alveo device(s) in your system. If you do not have an Alveo board, add --ignore-overlays
to the above command. The notebooks and overlays will be downloaded to the folder where the command was executed, and a Jupyter notebook server can be started from there.
There are several notebooks available to get you started with PyACCL:
- Intros to (Py)ACCL primitives and collectives
- Short guides to using compression and communicators
- Quick overview of performance-related flags
Running Tests
PyACCL includes tests for single ACCL instances and systems of ACCL instances. The tests are designed to run against ACCL emulator/simulator sessions as well as the single-FPGA ACCL test overlay, which connects 3 ACCL instances on a single Alveo board.
To run the tests, add the ACCL emulator executable cclo_emu
to your path and start your emulator/simulator session:
pyaccl-emulate -n <NRANKS>
then run the following command from the pyaccl root folder:
mpirun -np <NRANKS> python3 -m pytest --with-mpi
If your system has an Alveo board and NRANKS
is less or equal to 3, the test fixture will try and download the appropriate overlay for it, otherwise it will skip hardware testcases. Similarly, the test fixture will attempt to identify a valid emulator/simulator session with the appropriate number of ranks, then run testcases against it. If you omit --with-mpi
, only single-instance tests will run, i.e. tests for copy()
and combine()
primitives.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
File details
Details for the file pyaccl-0.2.tar.gz
.
File metadata
- Download URL: pyaccl-0.2.tar.gz
- Upload date:
- Size: 31.0 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.0 CPython/3.8.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 2ce2af9c02999358f0137e623c7159147fd2d0244c5cd9f6577cc48a1f0c26af |
|
MD5 | 7c608566f28c12dd1626a70fd4c314c0 |
|
BLAKE2b-256 | 43523970785c3d0f4479d552c5bdd54e2c6cb09578edbac195180247aa3c07b9 |