Automatic Model Parallelism Through FX #15

	name: Automatic Model Parallelism Test on GPUs

	on:
	pull_request:
	branches:
	- main
	paths:
	- 'optimum/fx/parallelization/**.py'
	push:
	branches:
	- main
	paths:
	- 'optimum/fx/parallelization/**.py'

	concurrency:
	group: ${{ github.workflow }}-${{ github.head_ref \|\| github.run_id }}
	cancel-in-progress: true

	jobs:
	run_gpu_tests:
	strategy:
	fail-fast: false
	matrix:
	config:
	- name: GPU-enabled Optimum Test Suite
	image: nvidia/cuda:12.4.1-devel-ubuntu22.04
	gpu_target: ["nvidia-multi-gpu-l4-runners", "nvidia-multi-gpu-a10-runners"]

	name: ${{ matrix.config.name }}
	runs-on:
	group: "${{matrix.gpu_target}}"

	container:
	image: ${{ matrix.config.image }}
	options: --mount type=tmpfs,destination=/tmp --shm-size 64gb --gpus all --ipc host -v /mnt/hf_cache:/mnt/cache/
	env:
	NCCL_DEBUG: INFO
	HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
	defaults:
	run:
	shell: bash

	steps:
	- uses: actions/setup-python@v5
	with:
	python-version: '3.10'

	- name: Checkout optimum
	uses: actions/checkout@v4
	with:
	fetch-depth: 1

	- name: Run nvidia-smi
	run: \|
	nvidia-smi

	- name: Install dependencies
	run: \|
	python3 -m pip install -U pip
	python3 -m pip install torch transformers
	python3 -m pip install .[tests]

	- name: Run automatic model parallelism tests
	run: \|
	pytest -s -v -o log_cli=true tests/fx/parallelization

Provide feedback