A faster and more memory-efficient implementation of `zero_to_fp32` #3404

	name: cpu-torch-latest

	on:
	workflow_dispatch:
	pull_request:
	paths-ignore:
	- 'docs/**'
	- 'blogs/**'
	- 'deepspeed/inference/v2/**'
	- 'tests/unit/inference/v2/**'
	merge_group:
	branches: [ master ]
	schedule:
	- cron: "0 0 * * *"

	concurrency:
	group: ${{ github.workflow }}-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	unit-tests:
	runs-on: ubuntu-24.04

	steps:
	- uses: actions/checkout@v4

	- id: setup-venv
	uses: ./.github/workflows/setup-venv

	- name: Install system packages
	run: \|
	sudo apt-get install -y numactl pdsh

	- name: Install pytorch
	run: \|
	pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
	python -c "import torch; print('torch:', torch.__version__, torch)"
	python -c "import torch; print('CUDA available:', torch.cuda.is_available())"

	- name: Install deepspeed
	run: \|
	pip install .[dev,autotuning]
	ds_report

	- name: Python environment
	run: \|
	pip list

	- name: Unit tests
	run: \|
	unset TORCH_CUDA_ARCH_LIST # only jit compile for current arch
	cd tests
	HF_HOME=/tmp/hf_home/ pytest $PYTEST_OPTS -n 4 unit/ --torch_ver="2.5"
	HF_HOME=/tmp/hf_home/ pytest $PYTEST_OPTS -m 'sequential' unit/ --torch_ver="2.5"

Provide feedback