[CICD]: Add 910C image and unified model management

.github/configs/ascend.yml

@@ -18,15 +18,11 @@
 platform: ascend
 
 # Docker image for this hardware
-ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:v0.2.0-ascend-ci
+ci_image: harbor.baai.ac.cn/flagscale/vllm-plugin-fl:ascend-vllm0.20.2-a3-ci
 
 # Runner labels for this hardware
 runner_labels:
-  - self-hosted
-  - Linux
-  - ARM64
-  - ascend
-  - npu-16
+  - flagcicd-910c
 
 # Container volumes (hardware-specific paths)
 container_volumes:

.github/configs/platforms.yml

@@ -23,7 +23,9 @@ platforms:
   cuda:
     enabled: true
   ascend:
-    enabled: false
+    # Ascend uses a scarce self-hosted NPU runner; keep it enabled when
+    # validating Ascend changes in PR CI.
+    enabled: true
   hygon:
     enabled: true
   metax:

.github/scripts/ascend/download_models.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+# Copyright 2026 FlagOS Contributors
+#
+# 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.
+
+# Provision models on the host-mounted /data volume. Images must not contain
+# model weights; every platform test refers to the same host path convention.
+set -euo pipefail
+
+export FL_MODEL_BASE_PATH="${FL_MODEL_BASE_PATH:-/data/models}"
+export HF_ENDPOINT="${HF_ENDPOINT:-https://hf-mirror.com}"
+
+QWEN_ROOT="${FL_MODEL_BASE_PATH}/Qwen"
+MODEL_ID="Qwen/Qwen3-0.6B"
+MODEL_DIR="${QWEN_ROOT}/Qwen3-0.6B"
+
+mkdir -p "${QWEN_ROOT}"
+
+if [[ -f "${MODEL_DIR}/config.json" ]]; then
+    echo "Model already available: ${MODEL_DIR}"
+    exit 0
+fi
+
+export MODEL_ID MODEL_DIR
+echo "Downloading ${MODEL_ID} from ${HF_ENDPOINT} to ${MODEL_DIR}"
+python - <<'PY'
+import os
+
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id=os.environ["MODEL_ID"],
+    local_dir=os.environ["MODEL_DIR"],
+    endpoint=os.environ["HF_ENDPOINT"],
+)
+PY
+
+test -f "${MODEL_DIR}/config.json"
+echo "Model ready: ${MODEL_DIR}"

.github/scripts/ascend/setup.sh

@@ -6,4 +6,19 @@ 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]"
+pip install \
+    --constraint requirements/ascend.txt \
+    --no-build-isolation \
+    --no-deps \
+    -e .
+
+python - <<'PY'
+import numpy
+
+expected = "1.26.4"
+if numpy.__version__ != expected:
+    raise RuntimeError(
+        f"Unexpected NumPy version: {numpy.__version__}; expected {expected}"
+    )
+print(f"NumPy version: {numpy.__version__}")
+PY

.github/workflows/_e2e_test.yml

@@ -93,6 +93,13 @@ jobs:
       - name: Check device availability
         run: bash .github/scripts/${{ inputs.platform }}/check.sh
 
+      - name: Prepare host models
+        run: |
+          model_script=".github/scripts/${{ inputs.platform }}/download_models.sh"
+          if [ -f "${model_script}" ]; then
+            bash "${model_script}"
+          fi
+
       - name: Install project
         run: bash .github/scripts/${{ inputs.platform }}/setup.sh
 

.github/workflows/ascend-manual.yml

@@ -0,0 +1,37 @@
+# Copyright 2026 FlagOS Contributors
+#
+# 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.
+
+# Ascend runners are scarce shared resources. Validate the image and selected
+# cases directly on the host first, then dispatch this workflow when a recorded
+# full CI result is needed.
+name: Ascend Manual CI
+
+on:
+  workflow_dispatch:
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: ascend-manual-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  test-ascend:
+    if: >-
+      github.event_name == 'workflow_dispatch' ||
+      github.event.label.name == 'run-ascend-ci'
+    uses: ./.github/workflows/_platform_test.yml
+    with:
+      platform: ascend
+    secrets: inherit

benchmarks/README.md

@@ -52,7 +52,7 @@ All arguments are passed directly to `vllm bench throughput`.
 Example:
 ```bash
 python benchmarks/benchmark_throughput_autotune.py \
-  --model /models/Qwen3-Next-80B-A3B-Instruct \
+  --model /data/models/Qwen/Qwen3-Next-80B-A3B-Instruct \
   --tensor-parallel-size 4 \
   --dataset-name random \
   --input-len 6144 \

benchmarks/benchmark_throughput_autotune.py

@@ -768,7 +768,7 @@ def main() -> None:
             "No throughput arguments provided. "
             "Pass them after the script args, e.g.: "
             'python benchmark_gems_autotune.py --ops "silu_and_mul" -- '
-            "--model /models/Qwen3-Next-80B-A3B-Instruct --tensor-parallel-size 4 ..."
+            "--model /data/models/Qwen/Qwen3-Next-80B-A3B-Instruct --tensor-parallel-size 4 ..."
         )
 
     if args.background and not os.environ.get("FLAGGEMS_AUTOTUNE_BACKGROUND"):

benchmarks/benchmark_throughput_serve.py

@@ -17,7 +17,7 @@
 
 # Usage:
 #  1. Start the server as follows (adjust model path and args as needed):
-# vllm serve /models/Qwen3.6-27B --tensor-parallel-size 2 --max-model-len 262144 --no-enable-log-requests --no-enable-prefix-caching
+# vllm serve /data/models/Qwen/Qwen3.6-27B --tensor-parallel-size 2 --max-model-len 262144 --no-enable-log-requests --no-enable-prefix-caching
 
 #  2. Run this benchmark script (default: 4 test cases):
 # python benchmarks/benchmark_throughput_serve.py
@@ -35,7 +35,7 @@
 from datetime import datetime
 from statistics import mean
 
-MODEL = "/models/Qwen3.6-27B"
+MODEL = "/data/models/Qwen/Qwen3.6-27B"
 
 # total runs for each case
 RUNS = 4

benchmarks/flagos_eval/run_benchmark.sh

@@ -17,7 +17,7 @@
 set -e
 
 # Arguments
-MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_benchmark.sh /workspace/Qwen3-4B/"}
+MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_benchmark.sh /data/models/Qwen/Qwen3-4B/"}
 
 # Output directory
 OUTPUT_DIR=bench_results

benchmarks/flagos_eval/run_eval.sh

@@ -17,8 +17,8 @@
 set -e
 
 # Arguments
-MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_eval.sh /workspace/Qwen3-4B/ hf_xxx"}
-HF_TOKEN=${2:?"Please provide HF_TOKEN, e.g.: ./run_eval.sh /workspace/Qwen3-4B/ hf_xxx"}
+MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_eval.sh /data/models/Qwen/Qwen3-4B/ hf_xxx"}
+HF_TOKEN=${2:?"Please provide HF_TOKEN, e.g.: ./run_eval.sh /data/models/Qwen/Qwen3-4B/ hf_xxx"}
 
 # Environment variables
 export HF_ENDPOINT=https://hf-mirror.com            # China mirror (can be removed for overseas)

benchmarks/flagos_eval/run_mixed_benchmark.sh

@@ -30,12 +30,12 @@ set -e
 #   seed: Random seed for shuffle (default: 42)
 #
 # Example:
-#   ./run_mixed_benchmark.sh /workspace/Qwen3-4B/
-#   ./run_mixed_benchmark.sh /workspace/Qwen3-4B/ 123
+#   ./run_mixed_benchmark.sh /data/models/Qwen/Qwen3-4B/
+#   ./run_mixed_benchmark.sh /data/models/Qwen/Qwen3-4B/ 123
 # ============================================================================
 
 # Arguments
-MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_mixed_benchmark.sh /workspace/Qwen3-4B/"}
+MODEL_PATH=${1:?"Please provide model path, e.g.: ./run_mixed_benchmark.sh /data/models/Qwen/Qwen3-4B/"}
 SEED=${2:-42}
 
 # Output directory

docker/ascend/Dockerfile

@@ -12,143 +12,58 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-ARG VLLM_VERSION=0.19.0
-ARG CANN_VERSION=8.5.1
-ARG CANN_CHIP=a3
-ARG UBUNTU_VERSION=22.04
-ARG PYTHON_VERSION=3.11
+# Keep the chip image aligned with the environment validated on Ascend 910C.
+ARG ASCEND_BASE_IMAGE=quay.io/ascend/vllm-ascend:v0.20.2rc1-a3
 
 # ---------- base stage ----------
-FROM quay.io/ascend/cann:${CANN_VERSION}-${CANN_CHIP}-ubuntu${UBUNTU_VERSION}-py${PYTHON_VERSION} AS base
+FROM ${ASCEND_BASE_IMAGE} AS base
 
-ARG MOONCAKE_TAG=v0.3.8.post1
-ARG SOC_VERSION=ascend910_9391
-ENV DEBIAN_FRONTEND=noninteractive
-ENV SOC_VERSION=${SOC_VERSION} \
-    TASK_QUEUE_ENABLE=1 \
-    OMP_NUM_THREADS=1
+ARG FLAGGEMS_VERSION=3e6528cf04f5f964a7b0fa6628de6f0410dbfd02
 
-RUN pip install --upgrade pip setuptools
+ENV DEBIAN_FRONTEND=noninteractive \
+    VLLM_PLUGINS=fl \
+    VLLM_FL_PLATFORM=ascend \
+    HF_ENDPOINT=https://hf-mirror.com \
+    FL_MODEL_BASE_PATH=/data/models
 
-# Add BiShengIR compiler to PATH
-ENV PATH="${ASCEND_TOOLKIT_HOME}/tools/bishengir/bin:${PATH}"
-
-# Install system deps (clang-15 for triton-ascend, jemalloc for memory alloc)
-RUN apt-get update -y && \
-    apt-get install -y git \
-        vim \
-        wget \
-        net-tools \
-        gcc \
-        g++ \
+RUN python -m pip install --upgrade pip "setuptools>=77.0.3" \
+    && python -m pip install --no-cache-dir \
+        scikit-build-core==0.11 \
+        pybind11 \
+        ninja \
         cmake \
-        numactl \
-        libnuma-dev \
-        libjemalloc2 \
-        clang-15 \
-    && update-alternatives --install /usr/bin/clang clang /usr/bin/clang-15 20 \
-    && update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-15 20 && \
-    rm -rf /var/cache/apt/* /var/lib/apt/lists/*
-
-# Build & install Mooncake (distributed KV cache)
-COPY ./scripts/mooncake_installer.sh /tmp/mooncake_installer.sh
-RUN git clone --depth 1 --branch ${MOONCAKE_TAG} https://github.com/kvcache-ai/Mooncake /tmp/Mooncake \
-    && mv /tmp/mooncake_installer.sh /tmp/Mooncake/ \
-    && cd /tmp/Mooncake \
-    && bash mooncake_installer.sh -y \
-    && ARCH=$(uname -m) \
-    && source /usr/local/Ascend/ascend-toolkit/set_env.sh \
-    && export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/${ARCH}-linux/devlib:/usr/local/Ascend/ascend-toolkit/latest/${ARCH}-linux/lib64:$LD_LIBRARY_PATH \
-    && mkdir -p build \
-    && cd build \
-    && cmake .. -DUSE_ASCEND_DIRECT=ON \
-    && make -j$(nproc) && make install \
-    && rm -rf /tmp/Mooncake
-
-# Install triton-ascend, ray, protobuf
-RUN pip install triton-ascend==3.2.0 \
-        'ray>=2.47.1,<=2.48.0' \
-        'protobuf>3.20.0' && \
-    pip cache purge
-
-# jemalloc preload + devlib LD_LIBRARY_PATH
-RUN echo "export LD_PRELOAD=/usr/lib/$(uname -m)-linux-gnu/libjemalloc.so.2:${LD_PRELOAD}" >> ~/.bashrc && \
-    echo "export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib" >> ~/.bashrc
+    && git clone https://github.com/flagos-ai/FlagGems.git /workspace/FlagGems \
+    && git -C /workspace/FlagGems checkout "${FLAGGEMS_VERSION}" \
+    && python -m pip install --no-build-isolation /workspace/FlagGems \
+    && python -m pip cache purge
 
 WORKDIR /workspace
 
 # ---------- dev stage ----------
 FROM base AS dev
 
-ARG VLLM_VERSION=0.19.0
-
-RUN pip install \
+RUN python -m pip install --no-cache-dir \
         pytest \
         pytest-cov \
         pytest-json-report \
         ruff \
-        pre-commit \
-        ninja \
-        cmake
-
-# Install vLLM (uninstall triton pulled by vllm on x86 — conflicts on Ascend)
-RUN VLLM_TARGET_DEVICE="empty" pip install "vllm[audio]==${VLLM_VERSION}" \
-        --extra-index-url https://download.pytorch.org/whl/cpu/ && \
-    pip uninstall -y triton && \
-    pip cache purge
+        pre-commit
 
 # ---------- ci stage ----------
-FROM base AS ci
+FROM dev AS ci
 
-ARG VLLM_VERSION=0.19.0
-
-# Install dev/test tools
-RUN pip install \
-        pytest \
-        pytest-cov \
+RUN python -m pip install --no-cache-dir \
         pytest-timeout \
-        pytest-json-report \
         pytest-metadata \
-        numpy \
+        numpy==1.26.4 \
         requests \
+        openai \
         decorator \
+        pyyaml \
         "modelscope>=1.18.1" \
-        ruff \
-        pre-commit \
-        ninja \
-        cmake
-
-# Install vLLM (uninstall triton pulled by vllm on x86 — conflicts on Ascend)
-RUN VLLM_TARGET_DEVICE="empty" pip install "vllm[audio]==${VLLM_VERSION}" \
-        --extra-index-url https://download.pytorch.org/whl/cpu/ && \
-    pip uninstall -y triton && \
-    pip cache purge
-
-# Install FlagGems (NPU backend)
-ARG FLAGGEMS_VERSION=v5.0.0
-RUN pip install -U scikit-build-core==0.11 pybind11 \
-    && 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
-
-# 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
+        "huggingface_hub>=0.34"
 
 # ---------- release stage ----------
 FROM base AS release
 
-ARG INDEX_URL
-ARG EXTRA_INDEX_URL
-ARG VLLM_VERSION=0.19.0
-
-# Install vLLM
-# Todo
-
 WORKDIR /workspace