Refactor optimizer implementations and improve multi_tensor ops #30

Workflow file for this run

.github/workflows/qa-l0-te-cpp-unittest-pytorch-lint.yml at fd48f82

	name: QA L0 - Core Unit & Lint Tests

	on:
	push:
	branches: main
	paths:
	- '.github/workflows/qa-l0-te-cpp-unittest-pytorch-lint.yml'
	- 'qa/L0_pytorch_lint/**'
	- 'transformer_engine/**'
	- 'tests/pytorch/**'
	pull_request:
	branches: main
	paths:
	- '.github/workflows/qa-l0-te-cpp-unittest-pytorch-lint.yml'
	- 'qa/L0_pytorch_lint/**'
	- 'transformer_engine/**'
	- 'tests/pytorch/**'

	concurrency:
	group: ${{ github.workflow }}-${{ github.event.pull_request.number \|\| github.ref }}-${{ github.actor }}
	cancel-in-progress: true

	jobs:
	run-qa-l0-core-tests:
	runs-on: [ self-hosted, Linux, X64, nvidia, gpu-8 ]
	defaults:
	run:
	shell: bash
	container:
	image: harbor.baai.ac.cn/flagscale/cuda12.8.1-torch2.7.1-python3.10-te2.9:20260209
	ports:
	- 80:80
	options: >-
	--gpus all
	--shm-size=500g
	--privileged
	--ipc=host
	--ulimit memlock=-1
	--ulimit stack=67108864
	--ulimit nofile=65535:65535
	--user root
	--pull always
	steps:
	- name: Checkout Code
	uses: actions/checkout@v6.0.1
	with:
	repository: ${{ github.event.pull_request.head.repo.full_name }}
	ref: ${{ github.event.pull_request.head.ref }}
	ssh-strict: true
	ssh-user: git
	persist-credentials: true
	clean: true
	sparse-checkout-cone-mode: true
	fetch-tags: false
	show-progress: true
	lfs: false
	submodules: recursive
	set-safe-directory: true

	- name: Install Dependencies & Build Transformer Engine
	# timeout-minutes: 40
	env:
	NVTE_FRAMEWORK: pytorch
	TE_WITH_NCCL: 1
	run: \|
	# Activate conda environment
	echo "=== Activating Conda Environment ==="
	source /opt/miniconda3/etc/profile.d/conda.sh
	conda activate flagscale-train

	# Install Python dependencies with version pinning
	echo "=== Installing Python Dependencies ==="
	pip install transformers expecttest

	# Build and install transformer_engine with verbose output
	echo "=== Building & Installing Transformer Engine ==="
	pip install --no-build-isolation -vvv . --no-deps

	# Verify TE installation with version check
	echo "=== Verifying Transformer Engine Installation ==="
	python3 tests/pytorch/test_sanity_import.py

	- name: Verify GPU Availability & Health
	run: \|
	# Execute GPU check
	echo "=== Checking GPU Status ==="
	source .github/workflows/scripts/gpu_check.sh
	wait_for_gpu

	# too heavy, disabled for now
	# - name: Run L0 C++ Unit Tests
	# # timeout-minutes: 60
	# env:
	# TE_PATH: .
	# run: \|
	# # Activate conda environment
	# source /opt/miniconda3/etc/profile.d/conda.sh
	# conda activate flagscale-train

	# # Get TE library paths with robust detection
	# TE_LIB_PATH=$(pip3 show transformer-engine \| grep -E "Location:\|Editable project location:" \| tail -n 1 \| awk '{print $NF}')
	# TE_CPP_LIB_PATH="${TE_LIB_PATH}/transformer_engine"

	# # Set environment variables for build
	# export CMAKE_PREFIX_PATH="${TE_CPP_LIB_PATH}:${CMAKE_PREFIX_PATH}"
	# export LD_LIBRARY_PATH="${TE_CPP_LIB_PATH}:${LD_LIBRARY_PATH}"
	# NUM_PHYSICAL_CORES=$(nproc)
	# NUM_PARALLEL_JOBS=$(nproc)

	# # Build and run C++ tests
	# cd $TE_PATH/tests/cpp
	# cmake -GNinja -Bbuild . -DTE_LIB_PATH="${TE_CPP_LIB_PATH}"
	# cmake --build build
	# export OMP_NUM_THREADS=$((NUM_PHYSICAL_CORES / NUM_PARALLEL_JOBS))

	# # Run C++ tests with verbose output
	# echo "=== Running C++ Unit Tests ==="
	# ctest --test-dir build -j$NUM_PARALLEL_JOBS

	- name: PyTorch C++ Lint
	# timeout-minutes: 5
	env:
	CPP_ONLY: 1
	TE_PATH: .
	run: \|
	# Activate conda environment
	source /opt/miniconda3/etc/profile.d/conda.sh
	conda activate flagscale-train

	# Run C++ lint checks
	echo "=== Running C++ Lint Checks ==="
	bash ./qa/L0_pytorch_lint/test.sh \|\| true

	echo ""
	echo "-----------------------------------------------------"
	echo "Note: Pylint check ignores errors C0411 (incorrect import position) and W0611 (unused import), which can be achieved by adding the parameter --disable=C0411,W0611"
	echo "-----------------------------------------------------"
	continue-on-error: true

	- name: PyTorch Python Lint
	# timeout-minutes: 5
	env:
	PYTHON_ONLY: 1
	TE_PATH: .
	run: \|
	# Activate conda environment
	source /opt/miniconda3/etc/profile.d/conda.sh
	conda activate flagscale-train

	# Run PyTorch lint checks
	echo "=== Running PyTorch Lint Checks ==="
	bash ./qa/L0_pytorch_lint/test.sh \|\| true

	echo ""
	echo "-----------------------------------------------------"
	echo "Note: Pylint check ignores errors C0411 (incorrect import position) and W0611 (unused import), which can be achieved by adding the parameter --disable=C0411,W0611"
	echo "-----------------------------------------------------"
	continue-on-error: true

	- name: Run L0 PyTorch Debug Unit Tests
	# timeout-minutes: 10
	env:
	TE_PATH: .
	run: \|
	# Activate conda environment
	source /opt/miniconda3/etc/profile.d/conda.sh
	conda activate flagscale-train

	# Run debug unit tests
	echo "=== Running L0 PyTorch Debug Unit Tests ==="
	bash ./qa/L0_pytorch_debug_unittest/test.sh

	- name: Run L0 PyTorch Core Unit Tests
	# timeout-minutes: 10
	env:
	TE_PATH: .
	TE_FL_PREFER: vendor
	run: \|
	# Activate conda environment
	source /opt/miniconda3/etc/profile.d/conda.sh
	conda activate flagscale-train

	export TE_LIB_PATH=$(python -c "import site; print(site.getsitepackages()[0])")/transformer_engine

	# Run core unit tests
	echo "=== Running L0 PyTorch Core Unit Tests ==="
	bash ./qa/L0_pytorch_unittest/test.sh

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Refactor optimizer implementations and improve multi_tensor ops #30

Workflow file

Refactor optimizer implementations and improve multi_tensor ops #30

Uh oh!

Workflow file for this run