Bump vLLM version to v0.18.0 in ci image

.github/configs/ascend.yml

@@ -4,7 +4,7 @@
 platform: ascend
 
 # Docker image for this hardware
-ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.1.0-ascend-ci
+ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.2.0-ascend-ci
 
 # Runner labels for this hardware
 runner_labels:

.github/configs/cuda.yml

@@ -4,7 +4,7 @@
 platform: cuda
 
 # Docker image for this hardware
-ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.1.0-cuda-ci
+ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.2.0-cuda-ci
 
 # Runner labels for this hardware
 runner_labels:
@@ -20,6 +20,7 @@ container_volumes:
 
 # Container options (hardware-specific settings)
 container_options: >-
+  --privileged
   --gpus all
   --shm-size=500g
   --hostname vllm-plugin-fl

.github/scripts/ascend/setup.sh

@@ -3,5 +3,7 @@
 # Setup script for Ascend NPU CI environment.
 set -euo pipefail
 
+git config --global --add safe.directory "$(pwd)"
+
 pip install --upgrade pip "setuptools>=77.0.3"
 pip install --no-build-isolation -e ".[test]"

.github/scripts/cuda/setup.sh

@@ -3,5 +3,7 @@
 # Setup script for CUDA CI environment.
 set -euo pipefail
 
+git config --global --add safe.directory "$(pwd)"
+
 uv pip install --system --upgrade pip
 uv pip install --system --no-build-isolation -e ".[test]"

.github/workflows/ci.yml

@@ -9,6 +9,8 @@ on:
     paths-ignore:
       - "**.md"
       - "docs/**"
+      - "examples/**"
+      - "docker/**"
       - "LICENSE"
       - ".github/ISSUE_TEMPLATE/**"
       - ".github/PULL_REQUEST_TEMPLATE.md"
@@ -17,6 +19,8 @@ on:
     paths-ignore:
       - "**.md"
       - "docs/**"
+      - "examples/**"
+      - "docker/**"
       - "LICENSE"
       - ".github/ISSUE_TEMPLATE/**"
       - ".github/PULL_REQUEST_TEMPLATE.md"

docker/ascend/Dockerfile

@@ -1,20 +1,12 @@
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 # ---------- base stage ----------
 FROM quay.io/ascend/vllm-ascend:v${VLLM_VERSION}rc1-a3 AS base
 
 RUN pip install --upgrade pip setuptools
 
-# CANN Toolkit environment variables (mirrors set_env.sh baked in at build time)
-ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
-ENV LD_LIBRARY_PATH="${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64:${LD_LIBRARY_PATH}" \
-    PYTHONPATH="${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}" \
-    PATH="${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${ASCEND_TOOLKIT_HOME}/tools/ccec_compiler/bin:${PATH}"
-
-# Set ATB environment variables
-ENV ATB_HOME_PATH=/usr/local/Ascend/nnal/atb/latest/atb/cxx_abi_1
-ENV LD_LIBRARY_PATH="${ATB_HOME_PATH}/lib:${ATB_HOME_PATH}/examples:${ATB_HOME_PATH}/tests/atbopstest:${LD_LIBRARY_PATH}" \
-    PATH="${ATB_HOME_PATH}/bin:${PATH}"
+# Add BiShengIR compiler to PATH
+ENV PATH="${ASCEND_TOOLKIT_HOME}/tools/bishengir/bin:${PATH}"
 
 # ---------- dev stage ----------
 FROM base AS dev
@@ -49,8 +41,9 @@ RUN pip install \
         cmake
 
 # Install FlagGems (NPU backend)
+ARG FLAGGEMS_VERSION=v5.0.0
 RUN pip install -U scikit-build-core==0.11 pybind11 \
-    && git clone https://github.com/flagos-ai/FlagGems /workspace/FlagGems \
+    && git clone --branch ${FLAGGEMS_VERSION} --depth 1 https://github.com/flagos-ai/FlagGems /workspace/FlagGems \
     && pip install --no-build-isolation \
         --config-settings=cmake.define.FLAGGEMS_BACKEND=NPU \
         /workspace/FlagGems
@@ -71,7 +64,7 @@ FROM base AS release
 
 ARG INDEX_URL
 ARG EXTRA_INDEX_URL
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 # Install vLLM
 # Todo

docker/ascend/Dockerfile.v0.1.0

@@ -0,0 +1,79 @@
+ARG VLLM_VERSION=0.13.0
+
+# ---------- base stage ----------
+FROM quay.io/ascend/vllm-ascend:v${VLLM_VERSION}rc1-a3 AS base
+
+RUN pip install --upgrade pip setuptools
+
+# CANN Toolkit environment variables (mirrors set_env.sh baked in at build time)
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LD_LIBRARY_PATH="${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64:${LD_LIBRARY_PATH}" \
+    PYTHONPATH="${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}" \
+    PATH="${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${ASCEND_TOOLKIT_HOME}/tools/ccec_compiler/bin:${PATH}"
+
+# Set ATB environment variables
+ENV ATB_HOME_PATH=/usr/local/Ascend/nnal/atb/latest/atb/cxx_abi_1
+ENV LD_LIBRARY_PATH="${ATB_HOME_PATH}/lib:${ATB_HOME_PATH}/examples:${ATB_HOME_PATH}/tests/atbopstest:${LD_LIBRARY_PATH}" \
+    PATH="${ATB_HOME_PATH}/bin:${PATH}"
+
+# ---------- dev stage ----------
+FROM base AS dev
+
+# Install dev tools
+RUN pip install \
+        pytest \
+        pytest-cov \
+        pytest-json-report \
+        ruff \
+        pre-commit \
+        ninja \
+        cmake
+
+# ---------- ci stage ----------
+FROM base AS ci
+
+# Install dev/test tools
+RUN pip install --upgrade pip
+RUN pip install \
+        pytest \
+        pytest-cov \
+        pytest-timeout \
+        pytest-json-report \
+        numpy \
+        requests \
+        decorator \
+        "modelscope>=1.18.1" \
+        ruff \
+        pre-commit \
+        ninja \
+        cmake
+
+# Install FlagGems (NPU backend)
+RUN pip install -U scikit-build-core==0.11 pybind11 \
+    && git clone https://github.com/flagos-ai/FlagGems /workspace/FlagGems \
+    && pip install --no-build-isolation \
+        --config-settings=cmake.define.FLAGGEMS_BACKEND=NPU \
+        /workspace/FlagGems
+
+# Install FlagTree
+RUN pip install flagtree==0.4.0+ascend3.2 \
+        --index-url=https://resource.flagos.net/repository/flagos-pypi-hosted/simple \
+        --trusted-host=resource.flagos.net
+
+# Set environment variables for vLLM and Triton
+ENV VLLM_PLUGINS=fl
+ENV TRITON_ALL_BLOCKS_PARALLEL=1
+
+WORKDIR /workspace
+
+# ---------- release stage ----------
+FROM base AS release
+
+ARG INDEX_URL
+ARG EXTRA_INDEX_URL
+ARG VLLM_VERSION=0.13.0
+
+# Install vLLM
+# Todo
+
+WORKDIR /workspace

docker/build.sh

@@ -14,12 +14,12 @@ PYTHON_VERSION="${PYTHON_VERSION:-3.12}"
 UV_VERSION="${UV_VERSION:-0.7.12}"
 CUDA_VERSION="${CUDA_VERSION:-12.8.1}"
 UBUNTU_VERSION="${UBUNTU_VERSION:-22.04}"
-VLLM_VERSION="${VLLM_VERSION:-0.13.0}"
+VLLM_VERSION="${VLLM_VERSION:-0.18.0}"
 
 # ---- Build options ----
 PLATFORM="${PLATFORM:-cuda}"
 TARGET="dev"
-IMAGE_NAME="localhost:5000/vllm-plugin-fl"
+IMAGE_NAME="harbor.baai.ac.cn/flagscale/vllm-plugin-fl"
 IMAGE_TAG=""
 INDEX_URL="${INDEX_URL:-}"
 EXTRA_INDEX_URL="${EXTRA_INDEX_URL:-}"

docker/cuda/Dockerfile

@@ -6,7 +6,7 @@ FROM nvcr.io/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} AS base
 
 ARG PYTHON_VERSION=3.12
 ARG UV_VERSION=0.7.12
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -47,7 +47,7 @@ FROM base AS dev
 
 ARG INDEX_URL
 ARG EXTRA_INDEX_URL
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 # Install vLLM
 RUN uv pip install --system \
@@ -70,7 +70,7 @@ FROM base AS ci
 
 ARG INDEX_URL
 ARG EXTRA_INDEX_URL
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 # Install vLLM
 RUN uv pip install --system \
@@ -92,18 +92,17 @@ RUN uv pip install --system \
         pre-commit \
         ninja \
         cmake
-
+ARG FLAGGEMS_VERSION=v5.0.0
 ARG FLAGCX_VERSION=v0.9.0
 
 # Install FlagGems
 RUN uv pip install --system scikit-build-core==0.11 pybind11 \
-    && git clone https://github.com/flagos-ai/FlagGems /workspace/FlagGems \
+    && git clone --branch ${FLAGGEMS_VERSION} --depth 1 https://github.com/flagos-ai/FlagGems /workspace/FlagGems \
     && uv pip install --system --no-build-isolation /workspace/FlagGems
 
 # Install FlagCX (NVIDIA)
-RUN git clone https://github.com/flagos-ai/FlagCX.git /workspace/FlagCX \
+RUN git clone --branch ${FLAGCX_VERSION} --depth 1  https://github.com/flagos-ai/FlagCX.git /workspace/FlagCX \
     && cd /workspace/FlagCX \
-    && git checkout ${FLAGCX_VERSION} \
     && git submodule update --init --recursive \
     && make USE_NVIDIA=1 \
     && cd plugin/torch \
@@ -125,7 +124,7 @@ FROM base AS release
 
 ARG INDEX_URL
 ARG EXTRA_INDEX_URL
-ARG VLLM_VERSION=0.13.0
+ARG VLLM_VERSION=0.18.0
 
 # Install vLLM
 RUN uv pip install --system \

pyproject.toml

@@ -52,7 +52,7 @@ test = [
     "requests",
     "openai",
     "decorator",
-    "vllm[audio]==0.13.0",
+    "vllm[audio]==0.18.0",
     "modelscope>=1.18.1",
 ]
 

tests/models/qwen3/next_tp8.yaml

@@ -3,10 +3,10 @@
 llm:
   model: "/data/models/Qwen/Qwen3-Next-80B-A3B-Instruct"
   tensor_parallel_size: 8
-  max_model_len: 16384
-  max_num_batched_tokens: 16384
+  max_model_len: 8192
+  max_num_batched_tokens: 8192
   max_num_seqs: 512
-  gpu_memory_utilization: 0.7
+  gpu_memory_utilization: 0.8
   enforce_eager: true
   trust_remote_code: true
 

tests/run.py

@@ -47,7 +47,8 @@
 _REPO_ROOT = Path(__file__).resolve().parent.parent
 sys.path.insert(0, str(_REPO_ROOT))
 
-from tests.utils.cleanup import device_cleanup
+from tests.utils.cleanup import device_cleanup, wait_for_memory
+from tests.utils.model_config import ModelConfig
 from tests.utils.platform_config import PlatformConfig
 from tests.utils.report import TestReport, TestResult
 
@@ -351,6 +352,24 @@ def _run_single(self, tc: TestCase) -> TestResult:
                 message="dry-run",
             )
 
+        # Wait for sufficient device memory before e2e tests
+        if tc.task in ("inference", "serving") and tc.model and tc.case:
+            gpu_util = ModelConfig.load(tc.model, tc.case).engine.get(
+                "gpu_memory_utilization", 0.9
+            )
+            ok, info = wait_for_memory(self.config.platform, gpu_util)
+            if not ok:
+                print("[run] FAILED: timed out waiting for device memory")
+                return TestResult(
+                    name=tc.name,
+                    passed=False,
+                    duration=0.0,
+                    message=f"OOM: timed out waiting for device memory\n{info}",
+                    task=tc.task,
+                    model=tc.model,
+                    case=tc.case,
+                )
+
         # Merge extra env vars (e.g. FL_TEST_MODEL/FL_TEST_CASE for inference)
         env = None
         if tc.extra_env: