8 GPU Feature Tests #16678
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: 8 GPU Feature Tests | |
| on: | |
| push: | |
| branches: [ main ] | |
| tags: | |
| - ciflow/8gpu/* | |
| paths-ignore: | |
| - 'torchtitan/experiments/**' | |
| pull_request: | |
| types: [opened, synchronize, reopened, ready_for_review] | |
| paths-ignore: | |
| - 'torchtitan/experiments/**' | |
| schedule: | |
| # Runs every 6 hours | |
| - cron: '0 */6 * * *' | |
| concurrency: | |
| group: unit-test${{ github.workflow }}-${{ github.ref == 'refs/heads/main' && github.run_number || github.ref }} | |
| cancel-in-progress: true | |
| defaults: | |
| run: | |
| shell: bash -l -eo pipefail {0} | |
| permissions: | |
| id-token: write | |
| contents: read | |
| jobs: | |
| # Step 1: Dynamically compute the matrix based on conditions | |
| set-matrix: | |
| # Skip scheduled runs on forks, where they would only fail and email the fork owner | |
| if: github.repository_owner == 'pytorch' || github.event_name != 'schedule' | |
| uses: ./.github/workflows/set-matrix.yaml | |
| # Step 2: Use the dynamic matrix in the build-test job | |
| build-test: | |
| needs: set-matrix | |
| uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main | |
| strategy: | |
| fail-fast: false | |
| matrix: ${{ fromJSON(needs.set-matrix.outputs.matrix) }} | |
| with: | |
| runner: ${{ matrix.runner }} | |
| gpu-arch-type: ${{ matrix.gpu-arch-type }} | |
| gpu-arch-version: ${{ matrix.gpu-arch-version }} | |
| docker-image: ${{ matrix.docker-image }} | |
| repository: pytorch/torchtitan | |
| upload-artifact: outputs | |
| timeout: 60 | |
| script: | | |
| set -eux | |
| # The generic Linux job chooses to use base env, not the one setup by the image | |
| CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]") | |
| conda activate "${CONDA_ENV}" | |
| # Log CUDA driver version for debugging. | |
| DRIVER_VERSION=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader | head -n 1 || true) | |
| echo "CUDA driver version: ${DRIVER_VERSION}" | |
| pip config --user set global.progress_bar off | |
| start=$(date +%s) | |
| TORCH_SPEC="torch" | |
| if [ -n "${{ matrix.torch-version }}" ]; then | |
| TORCH_SPEC="torch==${{ matrix.torch-version }}" | |
| fi | |
| python -m pip install --force-reinstall --pre \ | |
| "${TORCH_SPEC}" --index-url ${{ matrix.index-url }} | |
| # The torchcomms feature tests are currently disabled, so do not install | |
| # torchcomms in the main feature job. Its wheel pins torch exactly and | |
| # can force CI off the latest torch nightly. | |
| # if [[ "${{ matrix.gpu-arch-type }}" == "cuda" ]]; then | |
| # python -m pip install --pre torchcomms --index-url ${{ matrix.index-url }} | |
| # fi | |
| end=$(date +%s) | |
| echo "pip install torch took $((end - start)) seconds" | |
| if [[ "${{ matrix.gpu-arch-type }}" == "rocm" ]]; then | |
| export HIPBLASLT_TENSILE_LIBPATH="$(python -c 'import os, torch; print(os.path.join(os.path.dirname(torch.__file__), "lib", "hipblaslt", "library"))')" | |
| echo "HIPBLASLT_TENSILE_LIBPATH=${HIPBLASLT_TENSILE_LIBPATH}" | |
| fi | |
| start=$(date +%s) | |
| USE_CPP=0 python -m pip install --pre torchao --index-url ${{ matrix.index-url }} | |
| end=$(date +%s) | |
| echo "pip install torchao took $((end - start)) seconds" | |
| sudo mkdir -p "$RUNNER_TEMP/artifacts-to-be-uploaded" | |
| sudo chown -R $(id -u):$(id -g) "$RUNNER_TEMP/artifacts-to-be-uploaded" | |
| sudo mkdir -p "comm_traces" | |
| sudo chown -R $(id -u):$(id -g) "comm_traces" | |
| # Verify the accuracy first. | |
| echo "Checking FSDP8 v.s. HSDP (4, 2) accuracy parity" | |
| export baseline_options="--parallelism.data_parallel_replicate_degree=1" | |
| export test_options="--parallelism.data_parallel_replicate_degree=4" | |
| # Set architecture-specific parameters | |
| if [[ "${{ matrix.gpu-arch-type }}" == "cuda" ]]; then | |
| LOSS_FILE="tests/assets/losses/llama3_cuda.txt" | |
| STEPS=1 | |
| elif [[ "${{ matrix.gpu-arch-type }}" == "rocm" ]]; then | |
| # The loss results of FSDP and HSDP start to diverge after 5th | |
| # step when running with ROCm, we also need to adjust this. | |
| # But this is more an unknown issue that AMD people may want to | |
| # figure out the root cause or confirm that this is an expected | |
| # behavior. | |
| LOSS_FILE="tests/assets/losses/llama3_rocm_mi350x.txt" | |
| STEPS=1 | |
| else | |
| echo "Error: Unknown GPU architecture type: ${{ matrix.gpu-arch-type }}" | |
| exit 1 | |
| fi | |
| python3 scripts/loss_compare.py . . --baseline-options="${baseline_options}" --job-dump-folder="${RUNNER_TEMP}/artifacts-to-be-uploaded/accuracy_comparison_outputs" --export-result="${RUNNER_TEMP}/artifacts-to-be-uploaded/accuracy_comparison_outputs/result.txt" --steps=100 | |
| echo "Checking FSDP8 the first tep loss is the same as FSDP2HSDP4" | |
| python3 scripts/loss_compare.py . . --baseline-options="${baseline_options}" --test-options="${test_options}" --job-dump-folder="${RUNNER_TEMP}/artifacts-to-be-uploaded/accuracy_comparison_outputs" --assert-equal --steps=1 | |
| rm -rf $RUNNER_TEMP/artifacts-to-be-uploaded/* | |
| echo "Checking FSDP8 loss from a new run v.s. FSDP8 loss from text file parity" | |
| python3 scripts/loss_compare.py . . --baseline-options="${baseline_options}" --job-dump-folder="${RUNNER_TEMP}/artifacts-to-be-uploaded/accuracy_comparison_outputs" --import-result="${LOSS_FILE}" --assert-equal --steps=100 | |
| rm -rf $RUNNER_TEMP/artifacts-to-be-uploaded/* | |
| # MoE loss comparison: verify Qwen3 MoE FSDP+TP+EP deterministic losses match reference | |
| if [[ "${{ matrix.gpu-arch-type }}" == "cuda" ]]; then | |
| MOE_LOSS_FILE="tests/assets/losses/qwen3_moe_cuda.txt" | |
| elif [[ "${{ matrix.gpu-arch-type }}" == "rocm" ]]; then | |
| MOE_LOSS_FILE="tests/assets/losses/qwen3_moe_rocm_mi350x.txt" | |
| fi | |
| echo "Checking Qwen3 MoE FSDP+TP+EP loss parity against reference" | |
| python3 scripts/loss_compare.py . . \ | |
| --baseline-module=qwen3 --baseline-config=qwen3_moe_debug \ | |
| --baseline-options="--parallelism.tensor_parallel_degree 2 --parallelism.expert_parallel_degree 4" \ | |
| --test-options="--parallelism.tensor_parallel_degree 2 --parallelism.expert_parallel_degree 4" \ | |
| --job-dump-folder="${RUNNER_TEMP}/artifacts-to-be-uploaded/moe_loss_comparison" \ | |
| --import-result="${MOE_LOSS_FILE}" --assert-equal --steps=100 | |
| rm -rf $RUNNER_TEMP/artifacts-to-be-uploaded/* | |
| python -m tests.integration_tests.run_tests --gpu_arch_type ${{ matrix.gpu-arch-type }} --test_suite features $RUNNER_TEMP/artifacts-to-be-uploaded --ngpu 8 | |
| # Cleanup the checkpoints so that we don't waste network bandwidth and time. | |
| rm -rf $RUNNER_TEMP/artifacts-to-be-uploaded/*/checkpoint | |
| rm -rf artifacts-to-be-uploaded/*/checkpoint |