minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[CI] Restore SMG e2e on 2-gpu-h100 / 4-gpu-h100 runners (#24222)

Browse Source 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kangyan-Zhou
2026-05-01 23:55:20 -07:00
committed by GitHub
co-authored by Claude Opus 4.7
parent b939d5410f
commit 2e72a36420
16 changed files with 610 additions and 383 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/pr-test-rust.yml
+84 -79
View File
@@ -23,8 +23,6 @@ concurrency:
cancel-in-progress: true
env:
RUSTC_WRAPPER: sccache
SCCACHE_GHA_ENABLED: "true"
SGLANG_IS_IN_CI: true
jobs:
@@ -33,7 +31,18 @@ jobs:
github.event_name != 'pull_request' ||
(github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'run-ci')) ||
(github.event.action == 'labeled' && github.event.label.name == 'run-ci')
runs-on: 4-gpu-a10
# Pin to 22.04 so the wheel auditwheel-tags as manylinux_2_35; the
# self-hosted GPU runners are Ubuntu 22.04 (glibc 2.35) and reject
# manylinux_2_39 wheels produced on ubuntu-latest (Ubuntu 24.04).
runs-on: ubuntu-22.04
# sccache is only installed on the GitHub-hosted runners that run this
# job and `unit-tests`; setting RUSTC_WRAPPER workflow-wide leaks it to
# gateway-e2e on the self-hosted GPU runners (which don't have sccache),
# so any pip-install that compiles a Rust extension would fail with
# `could not execute process \`sccache rustc\``.
env:
RUSTC_WRAPPER: sccache
SCCACHE_GHA_ENABLED: "true"
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -121,6 +130,9 @@ jobs:
(github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'run-ci')) ||
(github.event.action == 'labeled' && github.event.label.name == 'run-ci')
runs-on: ubuntu-latest
env:
RUSTC_WRAPPER: sccache
SCCACHE_GHA_ENABLED: "true"
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -186,11 +198,32 @@ jobs:
github.event_name != 'pull_request' ||
(github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'run-ci')) ||
(github.event.action == 'labeled' && github.event.label.name == 'run-ci')
# The `responses` matrix entry is intentionally omitted. It needs
# `docker run gvenzl/oracle-xe` + `docker run shoofio/brave-search-mcp-sse`
# on the runner host, but the 2-/4-gpu-h100 runners are themselves
# containers without a Docker daemon. Re-enable by adding back:
# - name: responses
# runner: 2-gpu-h100
# timeout: 45
# test_dirs: "e2e_test/responses"
# extra_deps: ""
# env_vars: "SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1"
# reruns: "--reruns 2 --reruns-delay 5"
# setup_oracle: true
# setup_brave: true
# parallel_opts: ""
# plus the Oracle Instant Client / `gvenzl/oracle-xe` /
# `shoofio/brave-search-mcp-sse` setup + cleanup steps (see commit
# cf346bb15 for the exact step bodies) once a runner with
# `docker.sock` (or binary-installed deps) is available.
strategy:
fail-fast: false
matrix:
include:
- name: benchmarks
# 4 GPUs: test_pd_perf.py uses workers(prefill=2, decode=2) and
# test_regular_perf.py uses workers(count=4) — both need tp*workers=4.
runner: 4-gpu-h100
timeout: 32
test_dirs: "e2e_test/benchmarks"
extra_deps: "genai-bench==0.0.3"
@@ -198,81 +231,56 @@ jobs:
reruns: ""
upload_benchmarks: true
parallel_opts: "" # No parallel for benchmarks (performance measurement)
- name: responses
- name: e2e
runner: 2-gpu-h100
timeout: 45
test_dirs: "e2e_test/responses"
test_dirs: "e2e_test/router e2e_test/embeddings"
extra_deps: ""
env_vars: "SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1"
reruns: "--reruns 2 --reruns-delay 5"
setup_oracle: true
setup_brave: true
parallel_opts: "" # Cloud backend tests not compatible with parallel execution
- name: e2e
timeout: 45
test_dirs: "e2e_test/router e2e_test/embeddings"
extra_deps: "pytest-parallel py" # py is required for pytest-parallel with newer pytest
env_vars: "SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1"
reruns: "--reruns 2 --reruns-delay 5"
parallel_opts: "--workers 1 --tests-per-worker 4" # Thread-based parallelism
# Run tests serially. pytest-parallel (unmaintained since 2019)
# has buggy fixture-finalize handling under thread dispatch:
# both class- and function-scoped fixture references leaked
# between tests, leaving model_pool instances pinned at
# _ref_count > 0 and deadlocking later tests that needed
# eviction (50+ min hangs). On a 2-GPU runner with 5 distinct
# model:mode combos in router+embeddings, the suite is
# eviction-bound anyway, so the parallel speedup was illusory.
parallel_opts: ""
- name: chat-completions
runner: 2-gpu-h100
timeout: 45
test_dirs: "e2e_test/chat_completions"
extra_deps: ""
env_vars: "SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1"
reruns: "--reruns 2 --reruns-delay 5"
parallel_opts: ""
runs-on: 4-gpu-a10
- name: chat-completions-4gpu
runner: 4-gpu-h100
timeout: 45
# qwen-30b (tp=4) tests can't fit on the 2-gpu-h100 matrix entries —
# they get skipped there by hooks.py. Run them here so coverage holds.
test_dirs: "e2e_test/chat_completions/test_enable_thinking.py"
extra_deps: ""
env_vars: "SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1"
reruns: "--reruns 2 --reruns-delay 5"
parallel_opts: ""
runs-on: ${{ matrix.runner }}
timeout-minutes: ${{ matrix.timeout }}
# Self-hosted GPU runners are scarce; serialize per hardware type so
# 2-gpu-h100 and 4-gpu-h100 each run one job at a time across all
# in-flight PRs. Queue rather than cancel — different refs shouldn't
# interrupt each other.
concurrency:
group: pr-test-rust-${{ matrix.runner }}
cancel-in-progress: false
steps:
- name: Checkout code
uses: actions/checkout@v4
- name: Install SGLang dependencies
run: |
sudo --preserve-env=PATH bash scripts/ci/cuda/ci_install_dependency.sh
- name: Setup Oracle Instant Client
if: matrix.setup_oracle
run: |
sudo apt-get install -y unzip
INSTANT_CLIENT_DIR="/home/ubuntu/instant-client"
INSTANT_CLIENT_ZIP="instantclient-basic-linux.x64-23.9.0.25.07.zip"
if [ ! -d "$INSTANT_CLIENT_DIR/instantclient_23_9" ]; then
echo "Downloading Oracle Instant Client..."
mkdir -p "$INSTANT_CLIENT_DIR"
cd "$INSTANT_CLIENT_DIR"
wget https://download.oracle.com/otn_software/linux/instantclient/2390000/$INSTANT_CLIENT_ZIP
unzip $INSTANT_CLIENT_ZIP
rm $INSTANT_CLIENT_ZIP
else
echo "Oracle Instant Client already exists, skipping download"
fi
echo "LD_LIBRARY_PATH=/home/ubuntu/instant-client/instantclient_23_9:\$LD_LIBRARY_PATH" >> $GITHUB_ENV
- name: Start Oracle Database
if: matrix.setup_oracle
run: |
docker run -d -p 1521:1521 -e ORACLE_PASSWORD=oracle --name oracle-db gvenzl/oracle-xe:21-slim
echo "Starting Oracle DB..."
# Export Oracle connection environment variables
echo "ATP_USER=system" >> $GITHUB_ENV
echo "ATP_PASSWORD=oracle" >> $GITHUB_ENV
echo "ATP_DSN=localhost:1521/XEPDB1" >> $GITHUB_ENV
- name: Start Brave MCP Server
if: matrix.setup_brave
run: |
docker run -d --rm \
-p 8001:8080 \
-e BRAVE_API_KEY \
--name brave-search-server \
shoofio/brave-search-mcp-sse:1.0.10
echo "Starting Brave MCP Server..."
sleep 2
curl -f --max-time 1 http://localhost:8001/sse > /dev/null 2>&1 && echo "Brave MCP Server is healthy!" || echo "Brave MCP Server responded"
bash scripts/ci/cuda/ci_install_dependency.sh
- name: Download wheel artifact
uses: actions/download-artifact@v4
@@ -289,14 +297,27 @@ jobs:
run: |
python3 -m pip install pytest pytest-rerunfailures httpx openai grpcio grpcio-health-checking numpy
if [ -n "${{ matrix.extra_deps }}" ]; then
python3 -m pip --no-cache-dir install --upgrade ${{ matrix.extra_deps }}
if echo "${{ matrix.extra_deps }}" | grep -q "genai-bench"; then
# genai-bench's transitive deps (oci/locust) pull
# transformers<5 which requires huggingface_hub<1.0 — that
# downgrades the 1.11.0 ci_install_dependency.sh settled on
# and breaks `kernels` (requires huggingface_hub>=1.3.0,<2.0).
# Install --no-deps and supply the runtime deps explicitly.
python3 -m pip --no-cache-dir install --no-deps ${{ matrix.extra_deps }}
python3 -m pip --no-cache-dir install \
locust click rich tenacity oci openpyxl gevent matplotlib
else
# Other extras with well-behaved transitive deps —
# normal --upgrade install path.
python3 -m pip --no-cache-dir install --upgrade ${{ matrix.extra_deps }}
fi
fi
- name: Run E2E tests
run: |
python3 python/sglang/cli/killall.py
cd sgl-model-gateway
${{ matrix.env_vars }} ROUTER_LOCAL_MODEL_PATH="/home/ubuntu/models" pytest ${{ matrix.reruns }} ${{ matrix.parallel_opts }} ${{ matrix.test_dirs }} -s -vv -o log_cli=true --log-cli-level=INFO
${{ matrix.env_vars }} pytest ${{ matrix.reruns }} ${{ matrix.parallel_opts }} ${{ matrix.test_dirs }} -s -vv -o log_cli=true --log-cli-level=INFO
- name: Upload benchmark results
if: matrix.upload_benchmarks && success()
@@ -305,18 +326,6 @@ jobs:
name: genai-bench-results-all-policies
path: sgl-model-gateway/benchmark_**/
- name: Cleanup Brave MCP Server
if: always() && matrix.setup_brave
run: |
docker stop brave-search-server || true
docker rm brave-search-server || true
- name: Cleanup Oracle Database
if: always() && matrix.setup_oracle
run: |
docker stop oracle-db || true
docker rm oracle-db || true
docker-build-test:
if: |
github.event_name != 'pull_request' ||
@@ -350,11 +359,7 @@ jobs:
summarize-benchmarks:
needs: gateway-e2e
runs-on: ubuntu-latest
# Disabled while e2e tests are skipped in
# sgl-model-gateway/e2e_test/fixtures/hooks.py — no benchmarks run, so the
# genai-bench-results-all-policies artifact is never produced and the
# download step would fail. Re-enable together with the skip removal.
if: false
if: success()
steps:
- name: Checkout code