[bugfix][torchair] fix wasted NPU memory buffer allocation for quantized deepseek with unquantized MTP layer #1265

Workflow file for this run

.github/workflows/nightly_benchmarks.yaml at 941ac7a

	#
	# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
	# This file is a part of the vllm-ascend project.
	#
	# 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.
	#

	name: 'Benchmarks / Performance'
	# This workflow runs nightly benchmarks for vllm-ascend.

	on:
	schedule:
	# Run benchmarks at 20:00 and 03:00 Beijing time (UTC+8)
	- cron: "0 12 * * *"
	- cron: "0 19 * * *"

	workflow_dispatch:
	# Allow manual triggering of the workflow

	pull_request:
	types: [ labeled ]

	# 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 1 job can runs on static-8-01-cards
	concurrency:
	group: static-8-01-cards
	cancel-in-progress: false

	jobs:
	test:
	if: ${{ contains(github.event.pull_request.labels..name, 'performance-test') && contains(github.event.pull_request.labels..name, 'ready-for-test') \|\| github.event_name == 'schedule' \|\| github.event_name == 'workflow_dispatch' }}

	name: Benchmarks/vLLM=${{ matrix.vllm_branch }}, vLLM-Ascend=${{ matrix.vllm_ascend_branch }}, use_v1=${{ matrix.vllm_use_v1 }}
	runs-on: 'linux-arm64-npu-static-8'
	strategy:
	matrix:
	include:
	- vllm_branch: v0.10.2
	vllm_ascend_branch: main
	vllm_use_v1: 1
	max-parallel: 1
	container:
	image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.2.rc1-910b-ubuntu22.04-py3.11
	volumes:
	- /usr/local/dcmi:/usr/local/dcmi
	- /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
	- /usr/local/Ascend/driver/:/usr/local/Ascend/driver/
	# Use self-host cache speed up pip and model download
	- /home/action/.cache:/github/home/.cache/
	options: >-
	--device /dev/davinci0
	--device /dev/davinci1
	--device /dev/davinci_manager
	--device /dev/devmm_svm
	--device /dev/hisi_hdc
	env:
	VLLM_USE_MODELSCOPE: True
	ES_OM_DOMAIN: ${{ secrets.ES_OM_DOMAIN }}
	ES_OM_AUTHORIZATION: ${{ secrets.ES_OM_AUTHORIZATION }}
	VLLM_USE_V1: ${{ matrix.vllm_use_v1 }}
	steps:
	- name: Check npu and CANN info
	run: \|
	npu-smi info
	cat /usr/local/Ascend/ascend-toolkit/latest/"$(uname -i)"-linux/ascend_toolkit_install.info

	- name: Config mirrors
	run: \|
	# keep using tuna's proxy since linux-arm64-npu-static-8 is in another region
	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

	- name: Install system dependencies
	run: \|
	apt-get update -y
	apt-get -y install git jq wget curl lsof gcc g++ cmake libnuma-dev

	- name: Config git
	run: \|
	git config --global --add safe.directory "$GITHUB_WORKSPACE"
	git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf https://github.com/

	- name: Checkout vllm-project/vllm-ascend repo
	uses: actions/checkout@v4
	with:
	fetch-depth: 0

	- name: Checkout vllm-project/vllm repo
	uses: actions/checkout@v4
	with:
	repository: vllm-project/vllm
	path: ./vllm-empty
	ref: ${{ matrix.vllm_branch }}

	- name: Install vllm-project/vllm from source
	working-directory: ./vllm-empty
	run: \|
	VLLM_TARGET_DEVICE=empty pip install -e .

	- name: Install vllm-project/vllm-ascend
	env:
	PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
	run: \|
	pip install -e .
	pip install -r benchmarks/requirements-bench.txt

	- name: Run current commit benchmarks
	if: github.event_name != 'schedule' && github.event_name != 'workflow_dispatch'
	run: \|
	# Sometimes we only want to run benchmarks on the current commit
	# This is useful for debugging or a release benchmark
	bash benchmarks/scripts/run-performance-benchmarks.sh
	# Convert the benchmark results to markdown format
	python3 benchmarks/scripts/convert_json_to_markdown.py

	- name: Generate step summary
	if: github.event_name != 'schedule' && github.event_name != 'workflow_dispatch'
	run: \|
	cat ./benchmarks/results/benchmark_results.md >> $GITHUB_STEP_SUMMARY

	- name: Upload benchmark artifacts
	if: github.event_name != 'schedule' && github.event_name != 'workflow_dispatch'
	uses: actions/upload-artifact@v4
	with:
	name: "benchmark-performance-${{ matrix.vllm_branch }}-${{ matrix.vllm_ascend_branch }}-report"
	path: ./benchmarks/results/benchmark_results.md
	if-no-files-found: warn
	retention-days: 90
	overwrite: true

	- name: Install elastic_tool
	if: github.event_name != 'pull_request'
	run: \|
	pip install escli-tool==0.2.3

	- name: Collect pr info from vllm-project/vllm-ascend
	if: github.event_name != 'pull_request'
	run: \|
	# Only get the pull request which may influences performance
	git log --pretty=format:"%H %s" -- '*/.py' ':!docs/' ':!tests/' ':!examples/' ':!benchmarks/' > commit_log.txt
	escli check commit_log.txt

	- name: Prepare benchmark script in advance
	if: github.event_name != 'pull_request'
	# This is for the benchmark iteration, which will change the benchmark scripts while checkouting each commit.
	# We need ensure the benchmark scripts always available.
	run: \|
	# Prepare the benchmark script in advance
	mkdir -p /github/home/benchmarks
	cp -r benchmarks/* /github/home/benchmarks/

	- name: Run benchmark iteration
	env:
	PIP_EXTRA_INDEX_URL: https://mirrors.huaweicloud.com/ascend/repos/pypi
	if: github.event_name != 'pull_request'
	run: \|
	while IFS= read -r line \|\| [[ -n "$line" ]]; do
	commit_id=${line%% *}
	commit_title=${line#* }

	git checkout $commit_id
	commit_time=$(git show -s --format=%cd $commit_hash --date=iso-strict)
	commit_time_no_tz=${commit_time::19}
	pip install -e .

	echo "------------------------"
	echo "commit_id: $commit_id"
	echo "commit_title: $commit_title"
	echo "commit_time: $commit_time_no_tz"
	echo "vllm branch: ${{ matrix.vllm_branch }}"
	echo "vllm-ascend branch: ${{ matrix.vllm_ascend_branch }}"
	echo "------------------------"

	cd /github/home
	ERROR_MSG=""
	if ! bash benchmarks/scripts/run-performance-benchmarks.sh; then
	ERROR_MSG="Benchmark failed to run"
	fi
	# send the result to es
	escli add --vllm_branch ${{ matrix.vllm_branch }} \
	--vllm_ascend_branch ${{ matrix.vllm_ascend_branch }} \
	--commit_id $commit_id \
	--commit_title "$commit_title" \
	--created_at "$commit_time_no_tz" \
	--res_dir ./benchmarks/results \
	--error "$ERROR_MSG" \
	--extra_feat '{"VLLM_USE_V1": "${{ matrix.vllm_use_v1 }}"}'
	rm -rf ./benchmarks/results
	cd -
	done < commit_log.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[bugfix][torchair] fix wasted NPU memory buffer allocation for quantized deepseek with unquantized MTP layer #1265

Workflow file

[bugfix][torchair] fix wasted NPU memory buffer allocation for quantized deepseek with unquantized MTP layer #1265

Uh oh!

Jobs

Run details

Workflow file for this run