@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
yml 95 lines 2.91 KB
Raw
1 name: macOS Apple Silicon Smoke Test
2
3 on:
4 schedule:
5 # Daily at 2:30 AM UTC
6 - cron: '30 2 * * *'
7 workflow_dispatch: # Manual trigger
8
9 permissions:
10 contents: read
11
12 jobs:
13 macos-m1-smoke-test:
14 if: github.repository_owner == 'vllm-project'
15 # macos-26 (the supported target) is still a preview runner, so gate on GA
16 # macos-15 and keep macos-26 non-blocking.
17 strategy:
18 fail-fast: false
19 matrix:
20 include:
21 - os: macos-15
22 required: true
23 - os: macos-26
24 required: false
25 name: macos-m1-smoke-test (${{ matrix.os }})
26 runs-on: ${{ matrix.os }}
27 continue-on-error: ${{ !matrix.required }}
28 timeout-minutes: 30
29
30 steps:
31 - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
32
33 - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
34 with:
35 enable-cache: true
36 cache-dependency-glob: |
37 requirements/**/*.txt
38 pyproject.toml
39 python-version: '3.12'
40
41 - name: Create virtual environment
42 run: |
43 uv venv
44 echo "$GITHUB_WORKSPACE/.venv/bin" >> "$GITHUB_PATH"
45
46 - name: Install dependencies and build vLLM
47 run: |
48 uv pip install -r requirements/build/cpu.txt --index-strategy unsafe-best-match
49 uv pip install -r requirements/cpu.txt --index-strategy unsafe-best-match
50 uv pip install -e . --no-build-isolation
51 env:
52 CMAKE_BUILD_PARALLEL_LEVEL: 4
53
54 - name: Verify installation
55 run: |
56 python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
57
58 - name: Smoke test vllm serve
59 run: |
60 # Start server in background
61 VLLM_CPU_KVCACHE_SPACE=1 \
62 vllm serve Qwen/Qwen3-0.6B \
63 --max-model-len=2K \
64 --load-format=dummy \
65 --hf-overrides '{"num_hidden_layers": 2}' \
66 --enforce-eager \
67 --port 8000 &
68
69 SERVER_PID=$!
70
71 # Wait for server to start
72 for i in {1..30}; do
73 if curl -s http://localhost:8000/health > /dev/null; then
74 echo "Server started successfully"
75 break
76 fi
77 if [ "$i" -eq 30 ]; then
78 echo "Server failed to start"
79 kill "$SERVER_PID"
80 exit 1
81 fi
82 sleep 2
83 done
84
85 # Test health endpoint
86 curl -f http://localhost:8000/health
87
88 # Long prompt: hits the split-KV path that short prompts skip (#46769).
89 PAYLOAD=$(python -c "import json; print(json.dumps({'model': 'Qwen/Qwen3-0.6B', 'prompt': 'The quick brown fox jumps over the lazy dog. ' * 24, 'max_tokens': 16}))")
90 curl -f --max-time 120 http://localhost:8000/v1/completions \
91 -H "Content-Type: application/json" \
92 -d "$PAYLOAD"
93
94 # Cleanup
95 kill "$SERVER_PID"