96fa7ff63b
### What this PR does / why we need it? 1. Fix rank set in DP scenario. The new poc version of torch-npu support setting `ASCEND_RT_VISIBLE_DEVICES` dynamically, thus we could use the rank set in `DPEngineCoreProc` directly instead of calculating local rank across dp by hand in the patched `_init_data_parallel` Closes: https://github.com/vllm-project/vllm-ascend/issues/1170 2. Bump torch-npu version to 2.5.1.post1.dev20250528 Closes: https://github.com/vllm-project/vllm-ascend/pull/1242 Closes: https://github.com/vllm-project/vllm-ascend/issues/1232 ### How was this patch tested? CI passed with new added test. --------- Signed-off-by: MengqingCao <cmq0113@163.com> Signed-off-by: Icey <1790571317@qq.com> Co-authored-by: Icey <1790571317@qq.com>
257 lines
10 KiB
YAML
257 lines
10 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:
|
|
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
|
|
- v0.9.1
|
|
- v0.7.3
|
|
vllm-ascend-version:
|
|
description: 'vllm-ascend version:'
|
|
required: true
|
|
type: choice
|
|
options:
|
|
- main
|
|
- v0.7.3-dev
|
|
models:
|
|
description: 'model:'
|
|
required: true
|
|
type: choice
|
|
options:
|
|
- all
|
|
- Qwen/Qwen2.5-7B-Instruct
|
|
- Qwen/Qwen2.5-VL-7B-Instruct
|
|
- Qwen/Qwen3-8B-Base
|
|
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, 'dense-accuracy-test')) &&
|
|
contains(github.event.pull_request.labels.*.name, 'ready-for-test') ||
|
|
github.event_name == 'workflow_dispatch'
|
|
}}
|
|
runs-on: >-
|
|
${{
|
|
(matrix.model_name == 'Qwen/Qwen2.5-VL-7B-Instruct' && 'linux-arm64-npu-4') ||
|
|
'linux-arm64-npu-2'
|
|
}}
|
|
strategy:
|
|
matrix:
|
|
vllm_use_version: [0, 1]
|
|
# the accuracy test will run:
|
|
# 1. workflow_dispatch with models input
|
|
# - all: Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
|
|
# - specified but not all: Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-8B-Base
|
|
# 2. PR labeled with "*-accuracy-test"
|
|
# - accuracy-test: Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct
|
|
# - dense-accuracy-test: Qwen/Qwen2.5-7B-Instruct
|
|
# - vl-accuracy-test: Qwen/Qwen2.5-VL-7B-Instruct
|
|
model_name: ${{ fromJSON(
|
|
(github.event.inputs.models == 'all' &&
|
|
'["Qwen/Qwen2.5-7B-Instruct","Qwen/Qwen2.5-VL-7B-Instruct","Qwen/Qwen3-8B-Base"]') ||
|
|
(github.event.inputs.models == 'Qwen/Qwen2.5-7B-Instruct' &&
|
|
'["Qwen/Qwen2.5-7B-Instruct"]') ||
|
|
(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/Qwen2.5-7B-Instruct","Qwen/Qwen2.5-VL-7B-Instruct"]' ||
|
|
contains(github.event.pull_request.labels.*.name, 'dense-accuracy-test') &&
|
|
'["Qwen/Qwen2.5-7B-Instruct"]' ||
|
|
contains(github.event.pull_request.labels.*.name, 'vl-accuracy-test') &&
|
|
'["Qwen/Qwen2.5-VL-7B-Instruct"]'
|
|
) }}
|
|
# Remove exclude after https://github.com/vllm-project/vllm-ascend/issues/1044 resolved
|
|
exclude:
|
|
- model_name: Qwen/Qwen2.5-VL-7B-Instruct
|
|
vllm_use_version: 1
|
|
|
|
fail-fast: false
|
|
name: ${{ matrix.model_name }} accuracy V${{ matrix.vllm_use_version }}
|
|
container:
|
|
image: m.daocloud.io/quay.io/ascend/cann:8.1.rc1-910b-ubuntu22.04-py3.10
|
|
env:
|
|
HF_ENDPOINT: https://hf-mirror.com
|
|
HF_TOKEN: ${{ secrets.HF_TOKEN }}
|
|
DATASET_SOURCE: ModelScope
|
|
VLLM_USE_MODELSCOPE: True
|
|
# 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 -i 's|ports.ubuntu.com|mirrors.tuna.tsinghua.edu.cn|g' /etc/apt/sources.list
|
|
pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
|
|
apt-get update -y
|
|
apt install git -y
|
|
git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf https://github.com/
|
|
|
|
- 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 || 'v0.9.1' }}
|
|
|
|
- name: Install vllm-project/vllm from source
|
|
working-directory: ./vllm-empty
|
|
run: VLLM_TARGET_DEVICE=empty pip install -e .
|
|
|
|
- name: Checkout vllm-project/vllm-ascend repo
|
|
uses: actions/checkout@v4
|
|
with:
|
|
repository: vllm-project/vllm-ascend
|
|
path: ./vllm-ascend
|
|
ref: ${{ env.GHA_VLLM_ASCEND_VERSION }}
|
|
|
|
- name: Install vllm-project/vllm-ascend
|
|
working-directory: ./vllm-ascend
|
|
env:
|
|
PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
|
|
run: |
|
|
pip install -r requirements-dev.txt
|
|
pip install -e .
|
|
|
|
- name: Install lm-eval, ray, and datasets
|
|
run: |
|
|
pip install lm-eval
|
|
|
|
- name: Collect version info
|
|
run: |
|
|
for dir in /usr/local/Ascend/ascend-toolkit/*; do
|
|
dname=$(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="$2}' | sed 's/+.*//'
|
|
echo "GHA_VLLM_ASCEND_VERSION=${{ github.event.inputs.vllm-ascend-version || github.ref }}"
|
|
} >> "$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: ${{ env.GHA_VLLM_ASCEND_VERSION }}"
|
|
|
|
- name: Run Accuracy Test for V${{ matrix.vllm_use_version }}
|
|
id: report
|
|
working-directory: ./benchmarks
|
|
env:
|
|
PYTORCH_NPU_ALLOC_CONF: max_split_size_mb:256
|
|
VLLM_USE_V1: ${{ matrix.vllm_use_version }}
|
|
run: |
|
|
model_base_name=$(basename ${{ matrix.model_name }})
|
|
markdown_name="${model_base_name}-V${{ matrix.vllm_use_version }}"
|
|
echo "markdown_name=$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 }}"
|
|
|
|
- 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: Upload Report for V${{ matrix.vllm_use_version }}
|
|
if: ${{ github.event_name == 'workflow_dispatch' }}
|
|
uses: actions/upload-artifact@v4
|
|
with:
|
|
name: "${{ env.SAFE_VLLM_ASCEND_VERSION }}-${{ steps.report.outputs.markdown_name }}-report"
|
|
path: ./benchmarks/accuracy/${{ steps.report.outputs.markdown_name }}.md
|
|
if-no-files-found: warn
|
|
retention-days: 90
|
|
overwrite: true
|