f60bb474f9
### What this PR does / why we need it?
Currently our workflow run time takes about 3 hours in total, which
seriously affects the developer experience, so it is urgent to have a
optimization, after this pr, It is expected that the running time of the
full CI can be shortened to 1h40min.
- Enable linux-aarch64-a2 (64GB) to replace linux-arm64-npu (32GB)
- Change TP4 ---> TP2 * 2 max-parallel
- Move DeepSeek-V2-Lite-W8A8 to single card test
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.10.0
- vLLM main:
a2480251ec
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
406 lines
16 KiB
YAML
406 lines
16 KiB
YAML
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
|
|
# This test will be triggered:
|
|
# 1. PR labeled with: '*accuracy-test' (ONLY 1 label valid) & 'ready-for-test'
|
|
# 2. workflow_dispatch with models input
|
|
# See detail rule in strategy.matrix note
|
|
name: Benchmarks / accuracy
|
|
|
|
on:
|
|
schedule:
|
|
# Runs every 6 hours
|
|
- cron: '0 */6 * * *'
|
|
pull_request:
|
|
types: [ labeled ]
|
|
workflow_dispatch:
|
|
inputs:
|
|
vllm-version:
|
|
description: 'vllm version:'
|
|
required: true
|
|
type: choice
|
|
# Please also update this when bump matched version
|
|
# Current supported vLLM versions
|
|
options:
|
|
- main
|
|
- v0.10.0
|
|
- v0.9.1
|
|
- v0.7.3
|
|
vllm-ascend-version:
|
|
description: 'vllm-ascend version:'
|
|
required: true
|
|
type: choice
|
|
options:
|
|
- main
|
|
- v0.9.1-dev
|
|
- v0.7.3-dev
|
|
models:
|
|
description: 'model:'
|
|
required: true
|
|
type: choice
|
|
options:
|
|
- all
|
|
- Qwen/Qwen2.5-VL-7B-Instruct
|
|
- Qwen/Qwen3-8B-Base
|
|
- Qwen/Qwen3-30B-A3B
|
|
default: 'all'
|
|
|
|
# Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
|
|
# declared as "shell: bash -el {0}" on steps that need to be properly activated.
|
|
# It's used to activate ascend-toolkit environment variables.
|
|
defaults:
|
|
run:
|
|
shell: bash -el {0}
|
|
|
|
# only cancel in-progress runs of the same workflow
|
|
concurrency:
|
|
group: ${{ github.workflow }}-${{ github.ref }}
|
|
cancel-in-progress: true
|
|
|
|
jobs:
|
|
accuracy_tests:
|
|
# test will be triggered when tag '*-accuracy-test' & 'ready-for-test' or workflow_dispatch job
|
|
if: >-
|
|
${{
|
|
(contains(github.event.pull_request.labels.*.name, 'accuracy-test') ||
|
|
contains(github.event.pull_request.labels.*.name, 'vl-accuracy-test') ||
|
|
contains(github.event.pull_request.labels.*.name, 'moe-accuracy-test') ||
|
|
contains(github.event.pull_request.labels.*.name, 'dense-accuracy-test')) &&
|
|
contains(github.event.pull_request.labels.*.name, 'ready-for-test') ||
|
|
github.event_name == 'workflow_dispatch' || github.event_name == 'schedule'
|
|
}}
|
|
runs-on: >-
|
|
${{
|
|
(matrix.model_name == 'Qwen/Qwen3-30B-A3B' && 'linux-aarch64-a2-2') ||
|
|
'linux-aarch64-a2-1'
|
|
}}
|
|
strategy:
|
|
matrix:
|
|
# the accuracy test will run:
|
|
# 1. workflow_dispatch with models input
|
|
# - all: Qwen/Qwen3-30B-A3B, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
|
|
# - specified but not all: Qwen/Qwen3-30B-A3B, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
|
|
# 2. PR labeled with "*-accuracy-test"
|
|
# - accuracy-test: Qwen/Qwen3-8B-Base, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-30B-A3B
|
|
# - dense-accuracy-test: Qwen/Qwen3-8B-Base
|
|
# - vl-accuracy-test: Qwen/Qwen2.5-VL-7B-Instruct
|
|
# - moe-accuracy-test: Qwen/Qwen3-30B-A3B
|
|
model_name: ${{ fromJSON(
|
|
(github.event_name == 'schedule' &&
|
|
'["Qwen/Qwen3-30B-A3B","Qwen/Qwen2.5-VL-7B-Instruct","Qwen/Qwen3-8B-Base"]') ||
|
|
(github.event.inputs.models == 'all' &&
|
|
'["Qwen/Qwen3-30B-A3B","Qwen/Qwen2.5-VL-7B-Instruct","Qwen/Qwen3-8B-Base"]') ||
|
|
(github.event.inputs.models == 'Qwen/Qwen3-30B-A3B' &&
|
|
'["Qwen/Qwen3-30B-A3B"]') ||
|
|
(github.event.inputs.models == 'Qwen/Qwen2.5-VL-7B-Instruct' &&
|
|
'["Qwen/Qwen2.5-VL-7B-Instruct"]') ||
|
|
(github.event.inputs.models == 'Qwen/Qwen3-8B-Base' &&
|
|
'["Qwen/Qwen3-8B-Base"]') ||
|
|
contains(github.event.pull_request.labels.*.name, 'accuracy-test') &&
|
|
'["Qwen/Qwen3-8B-Base","Qwen/Qwen2.5-VL-7B-Instruct", "Qwen/Qwen3-30B-A3B"]' ||
|
|
contains(github.event.pull_request.labels.*.name, 'dense-accuracy-test') &&
|
|
'["Qwen/Qwen3-8B-Base"]' ||
|
|
contains(github.event.pull_request.labels.*.name, 'vl-accuracy-test') &&
|
|
'["Qwen/Qwen2.5-VL-7B-Instruct"]' ||
|
|
contains(github.event.pull_request.labels.*.name, 'moe-accuracy-test') &&
|
|
'["Qwen/Qwen3-30B-A3B"]'
|
|
) }}
|
|
|
|
fail-fast: false
|
|
name: ${{ matrix.model_name }} accuracy
|
|
container:
|
|
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.2.rc1-910b-ubuntu22.04-py3.11
|
|
env:
|
|
DATASET_SOURCE: ModelScope
|
|
VLLM_USE_MODELSCOPE: True
|
|
USE_MODELSCOPE_HUB: 1
|
|
# 1. If version specified (work_dispatch), do specified branch accuracy test
|
|
# 2. If no version (labeled PR), do accuracy test by default ref:
|
|
# The branch, tag or SHA to checkout. When checking out the repository that
|
|
# triggered a workflow, this defaults to the reference or SHA for that event.
|
|
# Otherwise, uses the default branch.
|
|
GHA_VLLM_ASCEND_VERSION: ${{ github.event.inputs.vllm-ascend-version }}
|
|
|
|
steps:
|
|
- name: Checkout repository
|
|
uses: actions/checkout@v4
|
|
|
|
- name: Check npu and CANN info
|
|
run: |
|
|
npu-smi info
|
|
cat /usr/local/Ascend/ascend-toolkit/latest/"$(uname -i)"-linux/ascend_toolkit_install.info
|
|
|
|
- name: Config mirrors
|
|
run: |
|
|
sed -Ei 's@(ports|archive).ubuntu.com@cache-service.nginx-pypi-cache.svc.cluster.local:8081@g' /etc/apt/sources.list
|
|
pip config set global.index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple
|
|
pip config set global.trusted-host cache-service.nginx-pypi-cache.svc.cluster.local
|
|
apt-get update -y
|
|
apt install git -y
|
|
|
|
- name: Install system dependencies
|
|
run: |
|
|
apt-get -y install `cat packages.txt`
|
|
apt-get -y install gcc g++ cmake libnuma-dev
|
|
|
|
- name: Checkout vllm-project/vllm repo
|
|
uses: actions/checkout@v4
|
|
with:
|
|
repository: vllm-project/vllm
|
|
path: ./vllm-empty
|
|
# Please also update this when bump matched version
|
|
ref: ${{ github.event.inputs.vllm-version || 'v0.10.0' }}
|
|
|
|
- name: Install vllm-project/vllm from source
|
|
working-directory: ./vllm-empty
|
|
run: VLLM_TARGET_DEVICE=empty pip install -e .
|
|
|
|
- name: Resolve vllm-ascend version
|
|
run: |
|
|
VERSION_INPUT="${{ github.event.inputs.vllm-ascend-version }}"
|
|
|
|
if [[ "$VERSION_INPUT" == "main" ]]; then
|
|
TAGS=$(git ls-remote --tags --sort=-v:refname https://github.com/vllm-project/vllm-ascend "v*" | cut -f2 | sed 's|refs/tags/||')
|
|
LATEST_TAG=$(echo "$TAGS" | head -n1)
|
|
if [[ -z "$LATEST_TAG" ]]; then
|
|
RESOLVED_VERSION="main"
|
|
else
|
|
RESOLVED_VERSION="$LATEST_TAG"
|
|
fi
|
|
else
|
|
RESOLVED_VERSION="$VERSION_INPUT"
|
|
fi
|
|
echo "GHA_VLLM_ASCEND_VERSION=$RESOLVED_VERSION" >> $GITHUB_ENV
|
|
|
|
- name: Checkout vllm-project/vllm-ascend repo
|
|
uses: actions/checkout@v4
|
|
with:
|
|
repository: vllm-project/vllm-ascend
|
|
path: ./vllm-ascend
|
|
ref: ${{ env.GHA_VLLM_ASCEND_VERSION }}
|
|
|
|
- name: Install vllm-project/vllm-ascend
|
|
working-directory: ./vllm-ascend
|
|
env:
|
|
PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
|
|
run: |
|
|
pip install -r requirements-dev.txt
|
|
pip install -v -e .
|
|
|
|
- name: Get vLLM commit hash and URL
|
|
working-directory: ./vllm-empty
|
|
run: |
|
|
VLLM_COMMIT=$(git rev-parse --short=7 HEAD)
|
|
echo "VLLM_COMMIT=$VLLM_COMMIT" >> $GITHUB_ENV
|
|
|
|
- name: Get vLLM-Ascend commit hash and URL
|
|
working-directory: ./vllm-ascend
|
|
run: |
|
|
VLLM_ASCEND_COMMIT=$(git rev-parse --short=7 HEAD)
|
|
echo "VLLM_ASCEND_COMMIT=$VLLM_ASCEND_COMMIT" >> $GITHUB_ENV
|
|
|
|
- name: Print resolved hashes
|
|
run: |
|
|
echo "vLLM : ${{ env.VLLM_COMMIT }}"
|
|
echo "vLLM-Ascend: ${{ env.VLLM_ASCEND_COMMIT }}"
|
|
|
|
- name: Install lm-eval, ray, and datasets
|
|
run: |
|
|
pip install lm-eval==0.4.8
|
|
|
|
- name: Collect version info
|
|
run: |
|
|
for dir in /usr/local/Ascend/ascend-toolkit/*; do
|
|
dname=$(basename "$dir")
|
|
if [ "$dname" != "latest" ]; then
|
|
TOOLKIT_DIR="$dname"
|
|
break
|
|
fi
|
|
done
|
|
INFO_FILE="/usr/local/Ascend/ascend-toolkit/${TOOLKIT_DIR}/$(uname -i)-linux/ascend_toolkit_install.info"
|
|
GHA_CANN_VERSION=$(grep "version=" "$INFO_FILE" \
|
|
| head -n1 \
|
|
| cut -d'=' -f2 \
|
|
| tr -d '"')
|
|
{
|
|
echo "GHA_CANN_VERSION=$GHA_CANN_VERSION"
|
|
pip show torch | grep "Version:" | awk '{print "GHA_TORCH_VERSION="$2}'
|
|
pip show torch_npu | grep "Version:" | awk '{print "GHA_TORCH_NPU_VERSION="$2}'
|
|
pip show vllm | grep "Version:" | awk '{print "GHA_VLLM_VERSION="$2}' | sed 's/+.*//'
|
|
} >> "$GITHUB_ENV"
|
|
|
|
- name: Print versions
|
|
run: |
|
|
echo "CANN: ${{ env.GHA_CANN_VERSION }}"
|
|
echo "Torch NPU: ${{ env.GHA_TORCH_NPU_VERSION }}"
|
|
echo "Torch: ${{ env.GHA_TORCH_VERSION }}"
|
|
echo "vLLM: ${{ env.GHA_VLLM_VERSION }}"
|
|
echo "vLLM Ascend: ${{ env.GHA_VLLM_ASCEND_VERSION }}"
|
|
|
|
- name: Run Accuracy Test
|
|
id: report
|
|
working-directory: ./benchmarks
|
|
env:
|
|
PYTORCH_NPU_ALLOC_CONF: max_split_size_mb:256
|
|
run: |
|
|
model_base_name=$(basename ${{ matrix.model_name }})
|
|
markdown_name="${model_base_name}"
|
|
echo "markdown_name=$markdown_name"
|
|
echo "markdown_name=$markdown_name" >> $GITHUB_OUTPUT
|
|
mkdir -p ./accuracy
|
|
|
|
python ./scripts/run_accuracy.py \
|
|
--model "${{ matrix.model_name }}" \
|
|
--output "./accuracy/${markdown_name}.md" \
|
|
--vllm_ascend_version "${{ env.GHA_VLLM_ASCEND_VERSION || github.ref }}" \
|
|
--cann_version "${{ env.GHA_CANN_VERSION }}" \
|
|
--torch_npu_version "${{ env.GHA_TORCH_NPU_VERSION }}" \
|
|
--torch_version "${{ env.GHA_TORCH_VERSION }}" \
|
|
--vllm_version "${{ env.GHA_VLLM_VERSION }}" \
|
|
--vllm_commit "${{ env.VLLM_COMMIT }}" \
|
|
--vllm_ascend_commit "${{ env.VLLM_ASCEND_COMMIT }}" \
|
|
|
|
- name: Generate step summary
|
|
if: ${{ always() }}
|
|
run: |
|
|
cat ./benchmarks/accuracy/${{ steps.report.outputs.markdown_name }}.md >> $GITHUB_STEP_SUMMARY
|
|
|
|
- name: Sanitize version string for artifact naming
|
|
run: |
|
|
SAFE_VLLM_ASCEND_VERSION="${GHA_VLLM_ASCEND_VERSION//\//-}"
|
|
echo "SAFE_VLLM_ASCEND_VERSION=$SAFE_VLLM_ASCEND_VERSION" >> "$GITHUB_ENV"
|
|
|
|
- name: Check report first line for failure
|
|
id: check_report
|
|
run: |
|
|
REPORT_PATH="./benchmarks/accuracy/${{ steps.report.outputs.markdown_name }}.md"
|
|
echo "Scanning $REPORT_PATH for ❌ …"
|
|
if grep -q '❌' "$REPORT_PATH"; then
|
|
echo "contains_fail=true" >> $GITHUB_OUTPUT
|
|
else
|
|
echo "contains_fail=false" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
- name: Upload Report
|
|
if: ${{ github.event_name == 'workflow_dispatch' && steps.check_report.outputs.contains_fail == 'false' }}
|
|
uses: actions/upload-artifact@v4
|
|
with:
|
|
name: "report-${{ env.SAFE_VLLM_ASCEND_VERSION }}-${{ steps.report.outputs.markdown_name }}"
|
|
path: ./benchmarks/accuracy/${{ steps.report.outputs.markdown_name }}.md
|
|
if-no-files-found: warn
|
|
retention-days: 90
|
|
overwrite: true
|
|
|
|
create_pr:
|
|
runs-on: ubuntu-latest
|
|
needs: accuracy_tests
|
|
if: ${{ github.event_name == 'workflow_dispatch' }}
|
|
env:
|
|
UPSTREAM_REPO: vllm-project/vllm-ascend
|
|
steps:
|
|
- name: Checkout repository
|
|
uses: actions/checkout@v4
|
|
with:
|
|
repository: vllm-ascend-ci/vllm-ascend
|
|
token: ${{ secrets.PAT_TOKEN }}
|
|
ref: main
|
|
|
|
- name: Add upstream remote
|
|
run: |
|
|
git remote add upstream https://github.com/${{ env.UPSTREAM_REPO }}.git
|
|
git fetch upstream
|
|
git remote -v
|
|
|
|
- name: Set Git user info dynamically
|
|
run: |
|
|
git config user.name "${{ github.actor }}"
|
|
git config user.email "${{ github.actor }}@users.noreply.github.com"
|
|
|
|
- name: Create or switch to branch
|
|
run: |
|
|
TIMESTAMP=$(date +%Y%m%d%H%M%S)
|
|
BRANCH_NAME="auto-pr/accuracy-report-${TIMESTAMP}"
|
|
echo "BRANCH_NAME=${BRANCH_NAME}" >> $GITHUB_ENV
|
|
git checkout -B "${BRANCH_NAME}" upstream/${{ github.event.inputs.vllm-ascend-version }}
|
|
|
|
- name: Download only current run reports
|
|
uses: actions/download-artifact@v4
|
|
with:
|
|
path: ./docs/source/developer_guide/evaluation/accuracy_report
|
|
pattern: report-*
|
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
|
run-id: ${{ github.run_id }}
|
|
|
|
- name: Delete old report
|
|
run: |
|
|
find ./docs/source/developer_guide/evaluation/accuracy_report -maxdepth 1 -type f -name '*.md' ! -name 'index.md' -delete
|
|
find ./docs/source/developer_guide/evaluation/accuracy_report -mindepth 2 -type f -name '*.md' -exec mv -f {} ./docs/source/developer_guide/evaluation/accuracy_report \;
|
|
find ./docs/source/developer_guide/evaluation/accuracy_report -mindepth 1 -type d -empty -delete
|
|
|
|
- name: Update accuracy_report/index.md
|
|
run: |
|
|
REPORT_DIR="./docs/source/developer_guide/evaluation/accuracy_report"
|
|
INDEX_MD="$REPORT_DIR/index.md"
|
|
{
|
|
echo "# Accuracy Report"
|
|
echo ""
|
|
echo ":::{toctree}"
|
|
echo ":caption: Accuracy Report"
|
|
echo ":maxdepth: 1"
|
|
|
|
for report in "$REPORT_DIR"/*.md; do
|
|
filename="$(basename "$report" .md)"
|
|
if [ "$filename" != "index" ]; then
|
|
echo "$filename"
|
|
fi
|
|
done
|
|
echo ":::"
|
|
} > "$INDEX_MD"
|
|
|
|
- name: push accuracy report
|
|
env:
|
|
GITHUB_TOKEN: ${{ secrets.PAT_TOKEN }}
|
|
run: |
|
|
git add ./docs/source/developer_guide/evaluation/accuracy_report/*.md
|
|
git commit -s -m "[Doc] Update accuracy reports for ${{ github.event.inputs.vllm-ascend-version }}"
|
|
git push -f origin "${{ env.BRANCH_NAME }}"
|
|
|
|
- name: Create PR in upstream via API
|
|
uses: actions/github-script@v7
|
|
with:
|
|
github-token: ${{ secrets.PAT_TOKEN }}
|
|
script: |
|
|
const pr = await github.rest.pulls.create({
|
|
owner: 'vllm-project',
|
|
repo: 'vllm-ascend',
|
|
head: `vllm-ascend-ci:${{ env.BRANCH_NAME }}`,
|
|
base: '${{ github.event.inputs.vllm-ascend-version }}',
|
|
title: `[Doc] Update accuracy reports for ${{ github.event.inputs.vllm-ascend-version }}`,
|
|
body: `The accuracy results running on NPU Altlas A2 have changed, updating reports for:
|
|
${{
|
|
github.event.inputs.models == 'all'
|
|
&& 'All models (Qwen/Qwen3-30B-A3B, Qwen2.5-VL-7B-Instruct, Qwen3-8B-Base)'
|
|
|| github.event.inputs.models
|
|
}}
|
|
|
|
- [Workflow run][1]
|
|
|
|
[1]: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}`
|
|
});
|
|
core.info(`Created PR #${pr.data.number}`);
|
|
|