Files
sglang/.github/workflows/pr-test-pd-router.yml
Shangming Cai 70e37b97bf chore: upgrade mooncake 0.3.5 (#8341)
Signed-off-by: Shangming Cai <caishangming@linux.alibaba.com>
2025-07-25 01:17:26 -07:00

370 lines
12 KiB
YAML

name: Test Disaggregation Mode
on:
push:
branches: [ main ]
paths:
- 'python/sglang/srt/disaggregation/**'
- 'scripts/ci_start_disaggregation_servers.sh'
- 'sgl-router/**'
pull_request:
branches: [ main ]
paths:
- 'python/sglang/srt/disaggregation/**'
- 'scripts/ci_start_disaggregation_servers.sh'
- 'sgl-router/**'
workflow_dispatch:
concurrency:
group: test-disaggregation-${{ github.ref }}
cancel-in-progress: true
permissions:
contents: read
pull-requests: write
issues: write
jobs:
test-disaggregation:
if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
runs-on: [h200]
timeout-minutes: 45
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
fetch-depth: 10
- name: Setup Python
uses: actions/setup-python@v4
with:
python-version: '3.11'
- name: Setup Rust
run: |
bash scripts/ci_install_rust.sh
- name: Cache Rust dependencies
uses: actions/cache@v4
with:
path: |
~/.cargo/bin/
~/.cargo/registry/index/
~/.cargo/registry/cache/
~/.cargo/git/db/
sgl-router/target/
key: ${{ runner.os }}-cargo-${{ hashFiles('sgl-router/Cargo.lock') }}
restore-keys: |
${{ runner.os }}-cargo-
- name: Cache pip dependencies
uses: actions/cache@v4
with:
path: ~/.cache/pip
key: ${{ runner.os }}-pip-${{ hashFiles('python/pyproject.toml') }}
restore-keys: |
${{ runner.os }}-pip-
- name: Validate environment
run: |
echo "=== System Validation ==="
nvidia-smi
echo "GPU count: $(nvidia-smi -L | wc -l)"
if [ $(nvidia-smi -L | wc -l) -lt 8 ]; then
echo "Error: This test requires at least 8 GPUs"
exit 1
fi
echo "=== RDMA Validation ==="
if ! command -v ibv_devices >/dev/null 2>&1; then
echo "Error: InfiniBand tools not found"
exit 1
fi
# Check for active IB devices
found_active_device=false
for device in mlx5_{0..11}; do
if ibv_devinfo $device >/dev/null 2>&1; then
state=$(ibv_devinfo $device | grep "state:" | head -1 | awk '{print $2}')
if [[ "$state" == "PORT_ACTIVE" ]]; then
echo "✓ Found active device: $device"
found_active_device=true
break
fi
fi
done
if [ "$found_active_device" = false ]; then
echo "Error: No active IB devices found"
echo "Available devices:"
ibv_devices || true
exit 1
fi
echo "=== Model Validation ==="
if [ ! -d "/raid/models/meta-llama/Llama-3.1-8B-Instruct" ]; then
echo "Error: Model not found"
ls -la /raid/models/ || echo "No models directory"
exit 1
fi
echo "✓ Model found"
- name: Install SGLang dependencies
run: |
echo "Installing SGLang with all extras..."
python3 -m pip --no-cache-dir install -e "python[all]" --break-system-packages
python3 -m pip --no-cache-dir install mooncake-transfer-engine==0.3.5
- name: Build and install sgl-router
run: |
source "$HOME/.cargo/env"
echo "Building sgl-router..."
cd sgl-router
cargo build && python3 -m build && pip install --force-reinstall dist/*.whl
- name: Start disaggregation servers
id: start_servers
run: |
echo "Starting disaggregation servers..."
bash scripts/ci_start_disaggregation_servers.sh &
SERVER_PID=$!
echo "server_pid=$SERVER_PID" >> $GITHUB_OUTPUT
# Wait for all 8 servers to be healthy (script already does this)
wait_count=0
while [ $wait_count -lt 30 ]; do
if ps -p $SERVER_PID > /dev/null; then
# Check if the startup script printed success message
sleep 2
wait_count=$((wait_count + 1))
else
# Script exited - check if it was successful
wait $SERVER_PID
exit_code=$?
if [ $exit_code -eq 0 ]; then
echo "✓ All disaggregation servers are healthy"
break
else
echo "Error: Server startup failed with code $exit_code"
exit 1
fi
fi
done
echo "✓ Servers started (PID: $SERVER_PID)"
- name: Test all policies sequentially
timeout-minutes: 30
run: |
POLICIES=("random" "round_robin" "cache_aware" "power_of_two")
BASE_URL="http://127.0.0.9:8000"
for policy in "${POLICIES[@]}"; do
echo ""
echo "=================================================="
echo "Testing policy: $policy"
echo "=================================================="
# Start router with the current policy
echo "Starting router with policy: $policy..."
python3 -m sglang_router.launch_router \
--pd-disaggregation \
--policy "$policy" \
--prefill http://127.0.0.1:30001 9001 \
--prefill http://127.0.0.2:30002 9002 \
--prefill http://127.0.0.3:30003 9003 \
--prefill http://127.0.0.4:30004 9004 \
--decode http://127.0.0.5:30005 \
--decode http://127.0.0.6:30006 \
--decode http://127.0.0.7:30007 \
--decode http://127.0.0.8:30008 \
--host 127.0.0.9 \
--port 8000 &
ROUTER_PID=$!
# Wait for router to become healthy
echo "Waiting for router to become healthy..."
TIMEOUT=60
ELAPSED=0
while [ $ELAPSED -lt $TIMEOUT ]; do
if curl --connect-timeout 5 --silent http://127.0.0.9:8000 > /dev/null 2>&1; then
echo "✓ Router is reachable"
break
fi
if ! ps -p $ROUTER_PID > /dev/null; then
echo "Error: Router process died"
exit 1
fi
sleep 5
ELAPSED=$((ELAPSED + 5))
done
if [ $ELAPSED -ge $TIMEOUT ]; then
echo "Error: Router health check timeout"
kill $ROUTER_PID 2>/dev/null || true
exit 1
fi
# Test API functionality
echo "Testing API completions for $policy..."
response=$(curl -s -X POST "$BASE_URL/v1/chat/completions" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer test-token" \
-d '{
"model": "/raid/models/meta-llama/Llama-3.1-8B-Instruct",
"messages": [
{"role": "user", "content": "Write a Python function to calculate fibonacci numbers recursively"}
],
"stream": false,
"max_tokens": 100
}')
if echo "$response" | jq -e '.choices[0].message.content' > /dev/null 2>&1; then
echo "✓ API test passed for $policy"
else
echo "✗ API test failed for $policy: $response"
kill $ROUTER_PID 2>/dev/null || true
exit 1
fi
# Test streaming
echo "Testing streaming API for $policy..."
stream_response=$(timeout 30 curl -s -X POST "$BASE_URL/v1/chat/completions" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer test-token" \
-d '{
"model": "/raid/models/meta-llama/Llama-3.1-8B-Instruct",
"messages": [
{"role": "user", "content": "Count from 1 to 5"}
],
"stream": true,
"max_tokens": 50
}')
if echo "$stream_response" | grep -q "data:"; then
echo "✓ Streaming API test passed for $policy"
else
echo "✗ Streaming API test failed for $policy"
kill $ROUTER_PID 2>/dev/null || true
exit 1
fi
# Run benchmark
echo "Running benchmark for $policy..."
benchmark_output=$(python3 -m sglang.bench_one_batch_server \
--model-path "/raid/models/meta-llama/Llama-3.1-8B-Instruct" \
--base-url "http://127.0.0.9:8000" \
--batch-size 8 \
--input-len 4096 \
--output-len 5 \
--skip-warmup)
echo "$benchmark_output"
# Save benchmark output
echo "$benchmark_output" > "benchmark_${policy}.txt"
# Extract and validate metrics
latency=$(echo "$benchmark_output" | grep "latency:" | awk '{print $2}' | sed 's/s//')
input_throughput=$(echo "$benchmark_output" | grep "input throughput:" | awk '{print $3}')
output_throughput=$(echo "$benchmark_output" | grep "output throughput:" | awk '{print $3}')
command -v bc >/dev/null || (apt-get update && apt-get install -y bc)
echo "Performance for $policy: ${latency}s | ${input_throughput} | ${output_throughput} tok/s"
# Validate performance
fail=""
(( $(echo "$latency > 1.5" | bc -l) )) && fail="Latency too high (${latency}s>1.5s) "
(( $(echo "$input_throughput < 20000" | bc -l) )) && fail="${fail}Input too low (${input_throughput}<20k) "
(( $(echo "$output_throughput < 1000" | bc -l) )) && fail="${fail}Output too low (${output_throughput}<1k) "
if [ -n "$fail" ]; then
echo "✗ Benchmark failed for $policy: $fail"
kill $ROUTER_PID 2>/dev/null || true
exit 1
else
echo "✓ Performance validation passed for $policy"
fi
# Stop router before testing next policy
echo "Stopping router for $policy..."
# First try graceful shutdown
kill $ROUTER_PID 2>/dev/null || true
# Wait up to 5 seconds for graceful shutdown
for i in {1..5}; do
if ! ps -p $ROUTER_PID > /dev/null 2>&1; then
echo "Router stopped gracefully"
break
fi
sleep 1
done
# Force kill if still running
if ps -p $ROUTER_PID > /dev/null 2>&1; then
echo "Force killing router..."
kill -9 $ROUTER_PID 2>/dev/null || true
fi
# Short delay to ensure port is released
sleep 2
echo "✓ Completed testing for $policy"
done
echo ""
echo "✅ All policies tested successfully!"
- name: Upload benchmark results
if: success()
uses: actions/upload-artifact@v4
with:
name: benchmark-results-all-policies
path: benchmark_*.txt
- name: Cleanup servers
if: always()
run: |
if [ -n "${{ steps.start_servers.outputs.server_pid }}" ]; then
pkill -P ${{ steps.start_servers.outputs.server_pid }} || true
kill ${{ steps.start_servers.outputs.server_pid }} || true
fi
pkill -f "sglang.launch_server" || true
sleep 5
remaining=$(ps aux | grep -c "sglang.launch_server" || echo "0")
echo "Cleanup completed. Remaining processes: $remaining"
summarize-benchmarks:
needs: test-disaggregation
runs-on: ubuntu-latest
if: success()
steps:
- name: Download benchmark results
uses: actions/download-artifact@v4
with:
name: benchmark-results-all-policies
- name: Create benchmark summary
run: |
echo "## PD Router Benchmark Results Summary" >> $GITHUB_STEP_SUMMARY
echo "" >> $GITHUB_STEP_SUMMARY
echo "| Policy | Latency (s) | Input Throughput (tok/s) | Output Throughput (tok/s) |" >> $GITHUB_STEP_SUMMARY
echo "|--------|-------------|-------------------------|--------------------------|" >> $GITHUB_STEP_SUMMARY
for policy in random round_robin cache_aware power_of_two; do
if [ -f "benchmark_${policy}.txt" ]; then
latency=$(grep "latency:" "benchmark_${policy}.txt" | awk '{print $2}')
input_throughput=$(grep "input throughput:" "benchmark_${policy}.txt" | awk '{print $3}')
output_throughput=$(grep "output throughput:" "benchmark_${policy}.txt" | awk '{print $3}')
echo "| ${policy} | ${latency} | ${input_throughput} | ${output_throughput} |" >> $GITHUB_STEP_SUMMARY
fi
done
echo "" >> $GITHUB_STEP_SUMMARY
echo "✅ All policies tested successfully!" >> $GITHUB_STEP_SUMMARY