merge master

2026-09-01 19:49:03 +02:00 · 2026-08-03 20:49:52 +08:00
parent 59dd3cda45 f0592cb67e
commit a9beacd5ea
24 changed files with 1804 additions and 225 deletions
--- a/docker/Dockerfile.amd
+++ b/docker/Dockerfile.amd
@@ -0,0 +1,26 @@
+FROM {base_image}
+
+ARG BASE_IMAGE_TAG={base_image_tag}
+LABEL modelscope.base_image="vllm/vllm-openai-rocm:${BASE_IMAGE_TAG}"
+
+# Build-time only (ARG does not persist into the image). Aliyun mirror can lag PyPI (~1h).
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple
+
+COPY docker/scripts/modelscope_env_init.sh /usr/local/bin/ms_env_init.sh
+
+ARG CUR_TIME={cur_time}
+RUN echo "CUR_TIME=${CUR_TIME}" && echo "BASE_IMAGE_TAG=${BASE_IMAGE_TAG}"
+
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    pip config set global.index-url https://mirrors.aliyun.com/pypi/simple && \
+    pip config set install.trusted-host mirrors.aliyun.com && \
+    cd /tmp && GIT_LFS_SKIP_SMUDGE=1 git clone -b {modelscope_branch} --single-branch https://github.com/modelscope/modelscope.git && \
+    cd modelscope && pip install --no-cache-dir . -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html && \
+    cd / && rm -fr /tmp/modelscope && pip cache purge
+
+ENV VLLM_USE_MODELSCOPE=True
+ENV LMDEPLOY_USE_MODELSCOPE=True
+ENV MODELSCOPE_CACHE=/mnt/workspace/.cache/modelscope/hub
+
+SHELL ["/bin/bash", "-c"]
--- a/docker/Dockerfile.ascend
+++ b/docker/Dockerfile.ascend
@@ -5,39 +5,54 @@ ENV PIP_DISABLE_PIP_VERSION_CHECK=1 \
    PIP_RETRIES=10 \
    SOC_VERSION={soc_version} \
    CANN_VERSION={cann_version}
+# Build-time only (ARG does not persist into the image).
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple

 SHELL ["/bin/bash", "-c"]

 # ---------- System dependencies ----------
-RUN rm -f /etc/apt/apt.conf.d/docker-clean && \
-    find /etc/apt/apt.conf.d -maxdepth 1 -type f | xargs -r grep -l "APT::Update::Post-Invoke\|docker-clean" | xargs -r rm -f && \
-    apt-get update -y && \
-    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-        gcc g++ cmake ninja-build libnuma-dev libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 \
-        wget git curl jq vim build-essential ca-certificates && \
-    apt-get clean && \
-    rm -rf /var/lib/apt/lists/*
+RUN set -eux; \
+    . /etc/os-release; \
+    case "${ID,,}" in \
+        ubuntu) \
+            rm -f /etc/apt/apt.conf.d/docker-clean; \
+            find /etc/apt/apt.conf.d -maxdepth 1 -type f | xargs -r grep -l "APT::Update::Post-Invoke\|docker-clean" | xargs -r rm -f; \
+            apt-get update -y; \
+            DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+                gcc g++ cmake ninja-build libnuma-dev libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 \
+                wget git curl jq vim build-essential ca-certificates; \
+            apt-get clean; \
+            rm -rf /var/lib/apt/lists/* \
+            ;; \
+        openeuler) \
+            yum install -y \
+                gcc gcc-c++ cmake ninja-build numactl-devel mesa-libGL glib2 libSM libXext libXrender \
+                wget git curl jq vim make ca-certificates; \
+            yum clean all; \
+            rm -rf /var/cache/yum \
+            ;; \
+        *) \
+            echo "Unsupported base image OS: ${ID}" >&2; \
+            exit 1 \
+            ;; \
+    esac

 RUN pip config set global.index-url https://mirrors.aliyun.com/pypi/simple && \
-    pip config set global.extra-index-url "https://pypi.org/simple" && \
-    pip config set install.trusted-host mirrors.aliyun.com && \
-    ARCH=$(uname -m) && \
-    if [ "$ARCH" = "x86_64" ]; then \
-        pip config set global.extra-index-url "https://pypi.org/simple https://download.pytorch.org/whl/cpu/"; \
-    fi
+    pip config set install.trusted-host mirrors.aliyun.com

 {extra_content}
 # ---------- Install vllm + vllm-ascend ----------
 RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    if [ -f /usr/local/Ascend/nnal/atb/set_env.sh ]; then source /usr/local/Ascend/nnal/atb/set_env.sh; fi && \
-    git clone --depth 1 --branch v0.18.0 https://github.com/vllm-project/vllm && \
-    git clone --depth 1 --branch v0.18.0 https://github.com/vllm-project/vllm-ascend.git
+    git clone --depth 1 --branch {vllm_git_ref} https://github.com/vllm-project/vllm && \
+    git clone --depth 1 --branch {vllm_ascend_git_ref} https://github.com/vllm-project/vllm-ascend.git

 RUN ARCH=$(uname -m) && \
+    export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    source /usr/local/Ascend/nnal/atb/set_env.sh && \
    # Install torch & torch_npu & torchvision
-    pip install torch==2.9.0 torch_npu==2.9.0.post2 torchvision==0.24.0 && \
+    pip install torch=={torch_version} torch_npu=={torch_npu_version} torchvision=={torchvision_version} && \
    # Install vllm
    cd vllm && VLLM_TARGET_DEVICE=empty pip install -v -e . && cd .. && \
    # Install vllm-ascend
@@ -46,32 +61,33 @@ RUN ARCH=$(uname -m) && \
 # ---------- Clone training-side repositories ----------
 RUN git clone --depth 1 --branch {megatron_branch} https://github.com/NVIDIA/Megatron-LM.git /Megatron-LM && \
    git clone --depth 1 --branch {mindspeed_branch} https://gitcode.com/Ascend/MindSpeed.git /MindSpeed && \
-    GIT_LFS_SKIP_SMUDGE=1 git clone --depth 1 -b {swift_branch} --single-branch https://github.com/modelscope/ms-swift.git /ms-swift && \
-    git clone --depth 1 https://github.com/modelscope/mcore-bridge.git /mcore-bridge
+    GIT_LFS_SKIP_SMUDGE=1 git clone --depth 1 -b {swift_branch} --single-branch https://github.com/modelscope/ms-swift.git /ms-swift

 # ---------- Install training-side repositories ----------
-RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    if [ -f /usr/local/Ascend/nnal/atb/set_env.sh ]; then source /usr/local/Ascend/nnal/atb/set_env.sh; fi && \
    cd /MindSpeed && pip install --no-cache-dir -e . && \
-    cd /mcore-bridge && pip install --no-cache-dir -e . && \
+    pip install --no-cache-dir mcore-bridge -i https://pypi.org/simple/ -U && \
    cd /ms-swift && pip install --no-cache-dir -e .

 # ---------- Pin torch to the correct version + torch_npu ----------
 # x86: must force-install the CPU build from pytorch.org/whl/cpu
 # aarch64: PyPI only provides the CPU build, so install it directly from the Aliyun mirror
-RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    if [ -f /usr/local/Ascend/nnal/atb/set_env.sh ]; then source /usr/local/Ascend/nnal/atb/set_env.sh; fi && \
    ARCH=$(uname -m) && \
    if [ "$ARCH" = "x86_64" ]; then \
        pip install --no-cache-dir --force-reinstall --no-deps \
            --index-url https://download.pytorch.org/whl/cpu \
-            torch==2.9.0 torchvision==0.24.0 torchaudio==2.9.0; \
+            torch=={torch_version} torchvision=={torchvision_version} torchaudio=={torchaudio_version}; \
    else \
        pip install --no-cache-dir --force-reinstall --no-deps \
-            torch==2.9.0 torchvision==0.24.0 torchaudio==2.9.0; \
+            torch=={torch_version} torchvision=={torchvision_version} torchaudio=={torchaudio_version}; \
    fi && \
    pip install --no-cache-dir --force-reinstall --no-deps \
-        torch_npu==2.9.0.post2 && \
+        torch_npu=={torch_npu_version} && \
    rm -rf /root/.cache/pip

 # ---------- Remove CUDA-only dependencies pulled in by vllm (they cause missing libtorch_cuda.so errors on NPU) ----------
@@ -83,7 +99,8 @@ ENV PYTHONPATH=/Megatron-LM:${PYTHONPATH}
 # install dependencies
 COPY requirements /var/modelscope

-RUN pip uninstall ms-swift modelscope -y && pip install --no-cache-dir pip==23.* -U && \
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    pip uninstall ms-swift modelscope -y && pip install --no-cache-dir pip==23.* -U && \
 if [ "$INSTALL_MS_DEPS" = "True" ]; then \
    pip install --no-cache-dir omegaconf==2.0.6 && \
    pip install 'editdistance==0.8.1' && \
@@ -109,9 +126,11 @@ fi
 ARG CUR_TIME={cur_time}
 RUN echo $CUR_TIME

-RUN pip install --no-cache-dir --no-build-isolation OpenCC
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    pip install --no-cache-dir --no-build-isolation OpenCC

-RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    if [ -f /usr/local/Ascend/nnal/atb/set_env.sh ]; then source /usr/local/Ascend/nnal/atb/set_env.sh; fi && \
    pip install --no-cache-dir -U funasr scikit-learn && \
    pip install --no-cache-dir -U qwen_vl_utils qwen_omni_utils librosa 'timm>=0.9.0' transformers accelerate peft trl safetensors && \
@@ -126,36 +145,20 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
    pip install --no-cache-dir omegaconf==2.3.0 && \
    pip cache purge

-# ---------- Reinstall triton-ascend for the selected CANN version ----------
+# ---------- Install training and evaluation dependencies ----------
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+    if [ -f /usr/local/Ascend/nnal/atb/set_env.sh ]; then source /usr/local/Ascend/nnal/atb/set_env.sh; fi && \
+    TORCH_DEVICE_BACKEND_AUTOLOAD=0 pip install --no-cache-dir "deepspeed<0.19" ray liger_kernel pre-commit -U && \
+    pip cache purge
+
+# ---------- Install triton-ascend ----------
 RUN set -eux; \
+    export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}"; \
    pip uninstall -y triton || true; \
    pip uninstall -y triton-ascend || true; \
-    case "${CANN_VERSION}" in \
-        8.5.*) \
-            pip install --no-cache-dir --force-reinstall triton-ascend==3.2.0; \
-            ;; \
-        9.0.0) \
-            PY_ABI="cp$(python -c 'import sys; print(f"{sys.version_info.major}{sys.version_info.minor}")')"; \
-            case "${PY_ABI}" in \
-                cp310|cp311|cp312|cp313) ;; \
-                *) echo "Unsupported Python ABI for triton-ascend 3.2.1: ${PY_ABI}" >&2; exit 1 ;; \
-            esac; \
-            ARCH="$(uname -m)"; \
-            case "${ARCH}" in \
-                aarch64|x86_64) ;; \
-                *) echo "Unsupported architecture for triton-ascend 3.2.1: ${ARCH}" >&2; exit 1 ;; \
-            esac; \
-            WHEEL_NAME="triton_ascend-3.2.1-${PY_ABI}-${PY_ABI}-manylinux_2_27_${ARCH}.manylinux_2_28_${ARCH}.whl"; \
-            WHEEL_PATH="/tmp/${WHEEL_NAME}"; \
-            curl -fL "https://gitcode.com/Ascend/triton-ascend/releases/download/v3.2.1/${WHEEL_NAME}" -o "${WHEEL_PATH}"; \
-            pip install --no-cache-dir --force-reinstall "${WHEEL_PATH}"; \
-            rm -f "${WHEEL_PATH}"; \
-            ;; \
-        *) \
-            echo "Unsupported CANN_VERSION for triton-ascend install: ${CANN_VERSION}" >&2; \
-            exit 1; \
-            ;; \
-    esac
+    pip install --no-cache-dir --force-reinstall \
+        triton-ascend=={triton_ascend_version} \
+        --extra-index-url=https://triton-ascend.osinfra.cn/pypi/simple

 RUN echo 'source /usr/local/Ascend/ascend-toolkit/set_env.sh' >> /root/.bashrc && \
    echo '[ -f /usr/local/Ascend/nnal/atb/set_env.sh ] && source /usr/local/Ascend/nnal/atb/set_env.sh' >> /root/.bashrc && \
--- a/docker/Dockerfile.ubuntu
+++ b/docker/Dockerfile.ubuntu
@@ -3,6 +3,8 @@ FROM {base_image}
 ARG DEBIAN_FRONTEND=noninteractive
 ENV TZ=Asia/Shanghai
 ENV arch=x86_64
+# Build-time only (ARG does not persist into the image). Aliyun mirror can lag PyPI (~1h).
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple

 COPY docker/scripts/modelscope_env_init.sh /usr/local/bin/ms_env_init.sh
 RUN apt-get update && \
@@ -21,7 +23,9 @@ ARG IMAGE_TYPE={image_type}
 # install dependencies
 COPY requirements /var/modelscope

-RUN pip uninstall ms-swift modelscope -y && pip --no-cache-dir install pip==23.* -U && \
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    pip uninstall ms-swift modelscope -y && pip --no-cache-dir install pip==23.* -U && \
 if [ "$INSTALL_MS_DEPS" = "True" ]; then \
    pip --no-cache-dir install omegaconf==2.0.6 && \
    pip install 'editdistance==0.8.1' && \
@@ -34,7 +38,7 @@ if [ "$INSTALL_MS_DEPS" = "True" ]; then \
    pip install --no-cache-dir 'scipy' && \
    pip install --no-cache-dir funtextprocessing typeguard==2.13.3 scikit-learn -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html && \
    pip install --no-cache-dir 'decord>=0.6.0' mpi4py paint_ldm ipykernel fasttext -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html && \
-    pip install --no-cache-dir ipywidgets && \
+    pip install --no-cache-dir ipywidgets jupyter_core nbconvert nbclient && \
    pip install --no-cache-dir 'blobfile>=1.0.5' && \
    pip uninstall MinDAEC -y && \
    pip install https://modelscope.oss-cn-beijing.aliyuncs.com/releases/dependencies/MinDAEC-0.0.2-py3-none-any.whl && \
@@ -47,7 +51,9 @@ fi
 ARG CUR_TIME={cur_time}
 RUN echo $CUR_TIME

-RUN bash /tmp/install.sh {version_args} && \
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    bash /tmp/install.sh {version_args} && \
    pip install --no-cache-dir -U funasr scikit-learn && \
    pip install --no-cache-dir -U qwen_vl_utils qwen_omni_utils librosa timm transformers accelerate peft trl safetensors && \
    cd /tmp && GIT_LFS_SKIP_SMUDGE=1 git clone -b {swift_branch}  --single-branch https://github.com/modelscope/ms-swift.git && \
@@ -63,23 +69,17 @@ RUN bash /tmp/install.sh {version_args} && \
    pip install --no-cache-dir transformers diffusers 'timm>=0.9.0' && pip cache purge; \
    pip install --no-cache-dir omegaconf==2.3.0 && pip cache purge; \
    pip config set global.index-url https://mirrors.aliyun.com/pypi/simple && \
-    pip config set global.extra-index-url https://pypi.org/simple && \
    pip config set install.trusted-host mirrors.aliyun.com && \
    cp /tmp/resources/ubuntu2204.aliyun /etc/apt/sources.list


-RUN if [ "$IMAGE_TYPE" = "gpu" ]; then \
-    pip install --no-cache-dir math_verify "datasets<4.8.5" "gradio<5.33" "deepspeed<0.19" ray -U && \
+ARG PIP_EXTRA_INDEX_URL=https://pypi.org/simple
+RUN export PIP_EXTRA_INDEX_URL="${PIP_EXTRA_INDEX_URL}" && \
+    if [ "$IMAGE_TYPE" = "gpu" ]; then \
+    pip install --no-cache-dir math_verify "gradio<5.33" "deepspeed<0.19" ray -U && \
    pip install --no-cache-dir mcore-bridge -i https://pypi.org/simple/ -U && \
-    pip install --no-cache-dir pybind11 liger_kernel wandb swanlab nvitop pre-commit "transformers<5.15" "trl<1.0" "peft<0.21" huggingface-hub -U && \
-    pip install git+https://github.com/NVIDIA/TransformerEngine.git@stable --no-build-isolation; \
-    pip install git+https://github.com/deepseek-ai/DeepGEMM.git@v2.1.1.post3 --no-build-isolation; \
-    pip install -U flash-linear-attention --no-build-isolation; \
-    pip install -U git+https://github.com/Dao-AILab/causal-conv1d --no-build-isolation; \
-    pip install git+https://github.com/Dao-AILab/fast-hadamard-transform --no-build-isolation; \
-    pip install git+https://github.com/NVIDIA-NeMo/Emerging-Optimizers.git@v0.3.0; \
-    mv /usr/local/lib/python3.12/site-packages/tilelang/lib/libcudart_stub.so /usr/local/lib/python3.12/site-packages/tilelang/lib/libcudart_stub.so.bak; \
-    ln -s /usr/local/cuda-13.0/targets/x86_64-linux/lib/libcudart.so /usr/local/lib/python3.12/site-packages/tilelang/lib/libcudart_stub.so; \
+    pip install --no-cache-dir liger_kernel wandb swanlab nvitop pre-commit "transformers" "trl<1.0" "peft<0.21" huggingface-hub -U && \
+    pip install --no-cache-dir --no-build-isolation "transformer_engine[pytorch]==2.16.0"; \
    cd /tmp && GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/NVIDIA/apex && \
    cd apex && pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./ && \
    cd / && rm -fr /tmp/apex && pip cache purge; \
--- a/docker/OVERVIEW.ascend.md
+++ b/docker/OVERVIEW.ascend.md
@@ -10,7 +10,8 @@ ms-swift Ascend images provide a ready-to-use ms-swift environment for Huawei As
 - Build template: `docker/Dockerfile.ascend`
 - Build entrypoint: `docker/build_image.py --image_type ascend`
 - Default base image: `quay.io/ascend/cann:8.5.1-a3-ubuntu22.04-py3.11`
- Default output tag: `${DOCKER_REGISTRY}:main-A3-py311-CANN8.5.1-ubuntu22.04-<arch>`
+- Supported base OSes: Ubuntu and openEuler, selected from the CANN base-image tag
+- Default output tag: `${DOCKER_REGISTRY}:main-cann8.5.1-torch_npu2.9.0.post2-a3-ubuntu22.04-py3.11-<arch>`
 - Ascend runtime environment is sourced from `/usr/local/Ascend/ascend-toolkit/set_env.sh`
 - If available, NNAL/ATB runtime is sourced from `/usr/local/Ascend/nnal/atb/set_env.sh`

@@ -22,45 +23,46 @@ The Ascend Dockerfile installs and configures:
 | --- | --- |
 | CANN | inherited from the selected `quay.io/ascend/cann` base image |
 | Python | inherited from the base image tag, for example `py3.11` |
-| PyTorch | `torch==2.9.0` |
-| torch-npu | `torch_npu==2.9.0.post2` |
-| torchvision / torchaudio | `torchvision==0.24.0`, `torchaudio==2.9.0` |
-| vLLM | source install from `vllm-project/vllm`, default branch `v0.18.0` |
-| vLLM Ascend | source install from `vllm-project/vllm-ascend`, default branch `v0.18.0` |
+| PyTorch | `torch==2.9.0` by default; configurable with `--torch_version` |
+| torch-npu | `torch_npu==2.9.0.post2` by default; configurable with `--torch_npu_version` |
+| torchvision / torchaudio | `torchvision==0.24.0`, `torchaudio==2.9.0` by default; pass both explicitly when overriding `--torch_version` |
+| vLLM | source install from `vllm-project/vllm`, default `0.18.0`; configurable with `--vllm_version` |
+| vLLM Ascend | source install from `vllm-project/vllm-ascend`, default `0.18.0`; configurable with `--vllm_ascend_version` |
 | Megatron-LM | source checkout, default branch `v0.15.3` |
 | MindSpeed | source checkout, default branch `core_r0.15.3` |
-| mcore-bridge | source checkout from `modelscope/mcore-bridge` |
+| mcore-bridge | latest release from PyPI |
 | ms-swift | source checkout from `modelscope/ms-swift`, default branch `main` |
 | ModelScope | source checkout from `modelscope/modelscope`, default branch `master` |
-| triton-ascend | `3.2.0` for CANN `8.5.*`; local wheel install of `3.2.1` for CANN `9.0.0` |
+| triton-ascend | CANN `8.5.*` defaults to `3.2.0`; CANN `9.0.*` defaults to `3.2.1`; configurable with `--triton_ascend_version` and installed from the Triton Ascend PyPI index |

 ## Supported Tag Format

 Images built by `docker/build_image.py --image_type ascend` use this tag format:

 ```text
-${DOCKER_REGISTRY}:<swift-branch>-<atlas-hardware>-<python-tag>-<cann-version-tag>-<os-tag>-<arch>
+${DOCKER_REGISTRY}:<swift-branch>-<cann-version-tag>-torch_npu<torch-npu-version>-<atlas-hardware>-<os-tag>-<python-tag>-<arch>
 ```

 | Field | Example | Description |
 | --- | --- | --- |
 | `swift-branch` | `main` | ms-swift branch used during image build |
-| `atlas-hardware` | `A2`, `A3`, `300I`, `A5` | Derived from `--soc_version` |
-| `python-tag` | `py311` | Derived from `--python_version` |
-| `cann-version-tag` | `CANN8.5.1`, `CANN9.0.0` | Parsed from the CANN base image tag |
-| `os-tag` | `ubuntu22.04` | Parsed from the CANN base image tag |
-| `arch` | `arm`, `x86` | Derived from host architecture or `--arch` |
+| `cann-version-tag` | `cann8.5.1`, `cann9.0.0` | Parsed from the CANN base image tag |
+| `torch-npu-version` | `2.9.0.post2` | From `--torch_npu_version`; defaults to `2.9.0.post2` |
+| `atlas-hardware` | `a2`, `a3`, `300i`, `a5` | Derived from `--soc_version` |
+| `os-tag` | `ubuntu22.04`, `openeuler24.03` | Parsed from the CANN base-image tag; prevents tags for different OSes from colliding |
+| `python-tag` | `py3.11` | Parsed from the CANN base image tag |
+| `arch` | `aarch64`, `x86_64` | Derived from host architecture or `--arch` |

 Default example on an ARM64 host:

 ```text
-${DOCKER_REGISTRY}:main-A3-py311-CANN8.5.1-ubuntu22.04-arm
+${DOCKER_REGISTRY}:main-cann8.5.1-torch_npu2.9.0.post2-a3-ubuntu22.04-py3.11-aarch64
 ```

 A2 / CANN 9.0.0 example:

 ```text
-${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm
+${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64
 ```

 ## Build Locally
@@ -85,6 +87,39 @@ python docker/build_image.py \
  --soc_version ascend910b1
 ```

+Build an openEuler image. The system-dependency layer automatically uses `yum`; Ubuntu images continue to use `apt-get`.
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --base_image quay.io/ascend/cann:8.5.1-a3-openeuler24.03-py3.11 \
+  --soc_version ascend910_9391
+```
+
+Override the PyTorch stack. `--torch_version` must match the base version of
+`--torch_npu_version`; when overriding PyTorch, pass its matching torchvision
+and torchaudio versions explicitly.
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --torch_version 2.9.0 \
+  --torch_npu_version 2.9.0.post2 \
+  --torchvision_version 0.24.0 \
+  --torchaudio_version 2.9.0
+```
+
+Override the vLLM stack or triton-ascend. The vLLM version arguments select
+the matching Git tag, for example `0.18.0` selects `v0.18.0`.
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --vllm_version 0.18.0 \
+  --vllm_ascend_version 0.18.0 \
+  --triton_ascend_version 3.2.1
+```
+
 Override Megatron or MindSpeed source branches when needed:

 ```bash
@@ -94,11 +129,11 @@ python docker/build_image.py \
  --mindspeed_branch core_r0.15.3
 ```

-For slow networks, Linux hosts can use Docker host networking after the root `Dockerfile` is generated:
+To run the rendered Dockerfile manually, use:

 ```bash
-docker build --network host \
-  -t ${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm \
+docker build \
+  -t ${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64 \
  -f Dockerfile .
 ```

@@ -119,7 +154,7 @@ docker run --rm -it \
  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
  -v /etc/ascend_install.info:/etc/ascend_install.info \
  -v /mnt/workspace:/mnt/workspace \
-  ${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm \
+  ${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64 \
  bash
 ```

@@ -147,7 +182,8 @@ pip show ms-swift modelscope torch-npu triton-ascend
 ## Notes

 - CANN, firmware, and driver versions must be compatible with each other.
- CANN `8.5.*` and CANN `9.0.0` use different `triton-ascend` install paths in this Dockerfile.
+- Ubuntu base images install system dependencies through `apt-get`; openEuler base images install the corresponding RPM packages through `yum`.
+- `triton-ascend` is installed from `https://triton-ascend.osinfra.cn/pypi/simple`; select a version compatible with the chosen CANN, Python, and architecture.
 - The image is intended for Ascend NPU ms-swift workflows. CUDA-only packages pulled in by dependencies are removed when they conflict with NPU runtime libraries.
 - Use a fixed image tag for production jobs instead of relying on a moving branch name.

--- a/docker/OVERVIEW.ascend.zh.md
+++ b/docker/OVERVIEW.ascend.zh.md
@@ -10,7 +10,8 @@ ms-swift Ascend 镜像面向华为昇腾 Atlas NPU，提供可直接使用的 ms
 - 构建模板：`docker/Dockerfile.ascend`
 - 构建入口：`docker/build_image.py --image_type ascend`
 - 默认基础镜像：`quay.io/ascend/cann:8.5.1-a3-ubuntu22.04-py3.11`
- 默认输出 tag：`${DOCKER_REGISTRY}:main-A3-py311-CANN8.5.1-ubuntu22.04-<arch>`
+- 支持的基础 OS：Ubuntu 和 openEuler，由 CANN 基础镜像 tag 选择
+- 默认输出 tag：`${DOCKER_REGISTRY}:main-cann8.5.1-torch_npu2.9.0.post2-a3-ubuntu22.04-py3.11-<arch>`
 - Ascend runtime 环境来自 `/usr/local/Ascend/ascend-toolkit/set_env.sh`
 - 如果镜像内存在 NNAL/ATB，则会加载 `/usr/local/Ascend/nnal/atb/set_env.sh`

@@ -22,45 +23,46 @@ Ascend Dockerfile 会安装和配置：
 | --- | --- |
 | CANN | 继承自选定的 `quay.io/ascend/cann` 基础镜像 |
 | Python | 继承自基础镜像 tag，例如 `py3.11` |
-| PyTorch | `torch==2.9.0` |
-| torch-npu | `torch_npu==2.9.0.post2` |
-| torchvision / torchaudio | `torchvision==0.24.0`，`torchaudio==2.9.0` |
-| vLLM | 从 `vllm-project/vllm` 源码安装，默认分支 `v0.18.0` |
-| vLLM Ascend | 从 `vllm-project/vllm-ascend` 源码安装，默认分支 `v0.18.0` |
+| PyTorch | 默认 `torch==2.9.0`；可通过 `--torch_version` 配置 |
+| torch-npu | 默认 `torch_npu==2.9.0.post2`；可通过 `--torch_npu_version` 配置 |
+| torchvision / torchaudio | 默认 `torchvision==0.24.0`、`torchaudio==2.9.0`；覆盖 `--torch_version` 时必须同时显式传入两者 |
+| vLLM | 从 `vllm-project/vllm` 源码安装，默认 `0.18.0`；可通过 `--vllm_version` 配置 |
+| vLLM Ascend | 从 `vllm-project/vllm-ascend` 源码安装，默认 `0.18.0`；可通过 `--vllm_ascend_version` 配置 |
 | Megatron-LM | 源码 checkout，默认分支 `v0.15.3` |
 | MindSpeed | 源码 checkout，默认分支 `core_r0.15.3` |
-| mcore-bridge | 来自 `modelscope/mcore-bridge` 的源码 checkout |
+| mcore-bridge | PyPI 上的最新发布版 |
 | ms-swift | 来自 `modelscope/ms-swift` 的源码 checkout，默认分支 `main` |
 | ModelScope | 来自 `modelscope/modelscope` 的源码 checkout，默认分支 `master` |
-| triton-ascend | CANN `8.5.*` 安装 `3.2.0`；CANN `9.0.0` 下载并本地安装 `3.2.1` wheel |
+| triton-ascend | CANN `8.5.*` 默认 `3.2.0`；CANN `9.0.*` 默认 `3.2.1`；可通过 `--triton_ascend_version` 配置，并从 Triton Ascend PyPI 源安装 |

 ## 支持的 Tag 格式

 通过 `docker/build_image.py --image_type ascend` 构建的镜像使用以下 tag 格式：

 ```text
-${DOCKER_REGISTRY}:<swift-branch>-<atlas-hardware>-<python-tag>-<cann-version-tag>-<os-tag>-<arch>
+${DOCKER_REGISTRY}:<swift-branch>-<cann-version-tag>-torch_npu<torch-npu-version>-<atlas-hardware>-<os-tag>-<python-tag>-<arch>
 ```

 | 字段 | 示例 | 说明 |
 | --- | --- | --- |
 | `swift-branch` | `main` | 构建镜像时使用的 ms-swift 分支 |
-| `atlas-hardware` | `A2`、`A3`、`300I`、`A5` | 从 `--soc_version` 推导 |
-| `python-tag` | `py311` | 从 `--python_version` 推导 |
-| `cann-version-tag` | `CANN8.5.1`、`CANN9.0.0` | 从 CANN 基础镜像 tag 解析 |
-| `os-tag` | `ubuntu22.04` | 从 CANN 基础镜像 tag 解析 |
-| `arch` | `arm`、`x86` | 从宿主机架构或 `--arch` 推导 |
+| `cann-version-tag` | `cann8.5.1`、`cann9.0.0` | 从 CANN 基础镜像 tag 解析 |
+| `torch-npu-version` | `2.9.0.post2` | 来自 `--torch_npu_version`，默认 `2.9.0.post2` |
+| `atlas-hardware` | `a2`、`a3`、`300i`、`a5` | 从 `--soc_version` 推导 |
+| `os-tag` | `ubuntu22.04`、`openeuler24.03` | 从 CANN 基础镜像 tag 解析；避免不同 OS 的镜像 tag 冲突 |
+| `python-tag` | `py3.11` | 从 CANN 基础镜像 tag 解析 |
+| `arch` | `aarch64`、`x86_64` | 从宿主机架构或 `--arch` 推导 |

 ARM64 宿主机上的默认示例：

 ```text
-${DOCKER_REGISTRY}:main-A3-py311-CANN8.5.1-ubuntu22.04-arm
+${DOCKER_REGISTRY}:main-cann8.5.1-torch_npu2.9.0.post2-a3-ubuntu22.04-py3.11-aarch64
 ```

 A2 / CANN 9.0.0 示例：

 ```text
-${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm
+${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64
 ```

 ## 本地构建
@@ -85,6 +87,36 @@ python docker/build_image.py \
  --soc_version ascend910b1
 ```

+构建 openEuler 镜像。系统依赖层会自动使用 `yum`，Ubuntu 镜像继续使用 `apt-get`：
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --base_image quay.io/ascend/cann:8.5.1-a3-openeuler24.03-py3.11 \
+  --soc_version ascend910_9391
+```
+
+覆盖 PyTorch 版本组。`--torch_version` 必须与 `--torch_npu_version` 的基础版本一致；覆盖 PyTorch 时，必须显式传入匹配的 torchvision 和 torchaudio 版本：
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --torch_version 2.9.0 \
+  --torch_npu_version 2.9.0.post2 \
+  --torchvision_version 0.24.0 \
+  --torchaudio_version 2.9.0
+```
+
+覆盖 vLLM 版本组或 triton-ascend。vLLM 版本参数会选择对应 Git tag，例如 `0.18.0` 会选择 `v0.18.0`。
+
+```bash
+python docker/build_image.py \
+  --image_type ascend \
+  --vllm_version 0.18.0 \
+  --vllm_ascend_version 0.18.0 \
+  --triton_ascend_version 3.2.1
+```
+
 需要时可以覆盖 Megatron 或 MindSpeed 源码分支：

 ```bash
@@ -94,11 +126,11 @@ python docker/build_image.py \
  --mindspeed_branch core_r0.15.3
 ```

-如果构建时网络较慢，Linux 宿主机可以在根目录 `Dockerfile` 生成后使用 host network 构建：
+如需手工构建生成后的根目录 `Dockerfile`，可使用：

 ```bash
-docker build --network host \
-  -t ${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm \
+docker build \
+  -t ${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64 \
  -f Dockerfile .
 ```

@@ -119,7 +151,7 @@ docker run --rm -it \
  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
  -v /etc/ascend_install.info:/etc/ascend_install.info \
  -v /mnt/workspace:/mnt/workspace \
-  ${DOCKER_REGISTRY}:main-A2-py311-CANN9.0.0-ubuntu22.04-arm \
+  ${DOCKER_REGISTRY}:main-cann9.0.0-torch_npu2.9.0.post2-a2-ubuntu22.04-py3.11-aarch64 \
  bash
 ```

@@ -147,7 +179,8 @@ pip show ms-swift modelscope torch-npu triton-ascend
 ## 注意事项

 - CANN、firmware 和 driver 版本必须互相兼容。
- 这个 Dockerfile 对 CANN `8.5.*` 和 CANN `9.0.0` 使用不同的 `triton-ascend` 安装路径。
+- Ubuntu 基础镜像通过 `apt-get` 安装系统依赖；openEuler 基础镜像通过 `yum` 安装对应 RPM 包。
+- `triton-ascend` 从 `https://triton-ascend.osinfra.cn/pypi/simple` 安装；请选择与 CANN、Python 和架构兼容的版本。
 - 该镜像面向 Ascend NPU 上的 ms-swift 工作流。依赖安装过程中引入且与 NPU runtime 冲突的 CUDA-only 包会被移除。
 - 生产任务建议使用固定镜像 tag，不要依赖浮动分支名。

--- a/docker/build_image.py
+++ b/docker/build_image.py
@@ -3,14 +3,27 @@ import os
 import platform
 import re
 import subprocess
+import urllib.error
+import urllib.request
 from copy import copy
 from datetime import datetime
-from typing import Any
+from typing import Any, List, Optional
+
+import json

 docker_registry = os.environ['DOCKER_REGISTRY']
 assert docker_registry, 'You must pass a valid DOCKER_REGISTRY'
 timestamp = datetime.now()
 formatted_time = timestamp.strftime('%Y%m%d%H%M%S')
+VLLM_ROCM_REPO = 'vllm/vllm-openai-rocm'
+_FLOATING_ROCM_TAGS = frozenset({
+    'latest',
+    'latest-base',
+    'nightly',
+    'base-nightly',
+})
+_VERSION_TAG_PATTERN = re.compile(r'^v\d+(?:\.\d+)*$')
+_NIGHTLY_HASH_PATTERN = re.compile(r'^(?:base-)?nightly-[0-9a-f]{7,40}$')


 class Builder:
@@ -359,7 +372,8 @@ class StableGPUImageBuilder(Builder):
            extra_content = extra_content.replace('{python_version}',
                                                  self.args.python_version)
            extra_content += """
-RUN pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
+RUN export PIP_EXTRA_INDEX_URL=https://pypi.org/simple && \
+    pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
 """
        version_args = (
            f'{self.args.torch_version} {self.args.torchvision_version} {self.args.torchaudio_version} '
@@ -434,7 +448,8 @@ class LatestGPUImageBuilder(StableGPUImageBuilder):
            extra_content = extra_content.replace('{python_version}',
                                                  self.args.python_version)
        extra_content += """
-RUN pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
+RUN export PIP_EXTRA_INDEX_URL=https://pypi.org/simple && \
+    pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
 """
        version_args = (
            f'{self.args.torch_version} {self.args.torchvision_version} {self.args.torchaudio_version} '
@@ -481,22 +496,399 @@ RUN pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
        return self.run_cmd('docker', 'push', image_tag2)


+class AmdImageBuilder(Builder):
+    """Build ModelScope image on top of vllm/vllm-openai-rocm."""
+
+    @staticmethod
+    def _is_specific_release_tag(tag: str) -> bool:
+        tag = tag.strip()
+        if not tag or tag.lower() in _FLOATING_ROCM_TAGS:
+            return False
+        if tag.endswith('-base'):
+            return False
+        if _NIGHTLY_HASH_PATTERN.fullmatch(tag):
+            return False
+        return bool(_VERSION_TAG_PATTERN.fullmatch(tag))
+
+    @staticmethod
+    def _image_digest(tag_info: dict) -> Optional[str]:
+        digest = tag_info.get('digest')
+        if digest:
+            return digest
+        for image in tag_info.get('images') or []:
+            digest = image.get('digest')
+            if digest:
+                return digest
+        return None
+
+    @classmethod
+    def _fetch_rocm_tags(cls, page_size: int = 100) -> List[dict]:
+        tags: List[dict] = []
+        url = (f'https://hub.docker.com/v2/repositories/{VLLM_ROCM_REPO}/tags'
+               f'?page_size={page_size}&ordering=-last_updated')
+        while url:
+            req = urllib.request.Request(
+                url, headers={'User-Agent': 'modelscope-docker-builder'})
+            try:
+                with urllib.request.urlopen(req, timeout=60) as resp:
+                    payload = json.load(resp)
+            except (urllib.error.URLError, json.JSONDecodeError) as exc:
+                raise RuntimeError(
+                    f'Failed to query Docker Hub tags for {VLLM_ROCM_REPO}: '
+                    f'{exc}') from exc
+            tags.extend(payload.get('results') or [])
+            url = payload.get('next')
+            # Only scan the first few pages; release tags are near the top.
+            if len(tags) >= 300:
+                break
+        if not tags:
+            raise RuntimeError(
+                f'No tags returned from Docker Hub for {VLLM_ROCM_REPO}')
+        return tags
+
+    @classmethod
+    def resolve_latest_rocm_tag(cls) -> str:
+        """Resolve the newest concrete release tag for vllm-openai-rocm.
+
+        Preference order:
+        1. Semver tag (vX.Y.Z) that shares digest with floating ``latest``
+        2. Newest semver tag by Docker Hub ``last_updated``
+        """
+        tags = cls._fetch_rocm_tags()
+        by_name = {item['name']: item for item in tags if item.get('name')}
+        release_tags = [
+            item for item in tags
+            if cls._is_specific_release_tag(item.get('name', ''))
+        ]
+        latest_info = by_name.get('latest')
+        latest_digest = cls._image_digest(latest_info) if latest_info else None
+        if latest_digest:
+            matched = [
+                item for item in release_tags
+                if cls._image_digest(item) == latest_digest
+            ]
+            if matched:
+                # Prefer the first match in last_updated order from API.
+                chosen = matched[0]['name']
+                print(
+                    f'Resolved {VLLM_ROCM_REPO} latest digest to release tag: '
+                    f'{chosen}')
+                return chosen
+
+        if not release_tags:
+            raise RuntimeError(
+                f'No concrete release tags found for {VLLM_ROCM_REPO}')
+        chosen = release_tags[0]['name']
+        print(f'Resolved newest {VLLM_ROCM_REPO} release tag: {chosen}')
+        return chosen
+
+    def init_args(self, args: Any) -> Any:
+        # Auto-discover from Docker Hub unless an explicit override is given.
+        override = getattr(args, 'base_image_tag', None)
+        if override and str(override).strip() and str(
+                override).strip().lower() not in {'auto', 'latest'}:
+            args.base_image_tag = str(override).strip()
+            if not self._is_specific_release_tag(args.base_image_tag):
+                raise ValueError(
+                    'base_image_tag override must be a concrete release tag '
+                    f'(e.g. v0.25.1), got: {args.base_image_tag}')
+            print(f'Using override AMD ROCm base image tag: '
+                  f'{args.base_image_tag}')
+        else:
+            args.base_image_tag = self.resolve_latest_rocm_tag()
+        if not args.base_image:
+            args.base_image = f'{VLLM_ROCM_REPO}:{args.base_image_tag}'
+        if not args.cuda_version:
+            args.cuda_version = '0.0.0'
+        return args
+
+    @staticmethod
+    def _sanitize_tag(tag: str) -> str:
+        return re.sub(r'[^A-Za-z0-9._-]+', '-', tag)
+
+    @staticmethod
+    def _normalize_version(version: str) -> str:
+        version = version.strip().lstrip('vV')
+        version = version.split('+')[0].split(' ')[0]
+        return re.sub(r'[^0-9A-Za-z._-]+', '', version)
+
+    @staticmethod
+    def _python_tag_from_version(version: str) -> str:
+        parts = version.strip().split('.')
+        if len(parts) >= 2 and parts[0].isdigit() and parts[1].isdigit():
+            return f'py{parts[0]}{parts[1]}'
+        return f'py{re.sub(r"[^0-9]", "", version)}'
+
+    @classmethod
+    def _run_capture(cls, *cmd: str) -> subprocess.CompletedProcess:
+        return subprocess.run(
+            list(cmd), capture_output=True, text=True, check=False)
+
+    @classmethod
+    def _probe_via_entrypoint(cls, base_image: str) -> dict:
+        """Read versions with docker run --entrypoint (no GPU required)."""
+        # Keep this script compact: it runs inside the base image via python -c.
+        script = (
+            'import json,os,pathlib,subprocess,sys\n'
+            'info={"python":"%d.%d.%d"%sys.version_info[:3]}\n'
+            'try:\n'
+            ' import torch\n'
+            ' info["torch"]=torch.__version__\n'
+            ' hip=getattr(torch.version,"hip",None)\n'
+            ' if hip: info["torch_hip"]=hip\n'
+            'except Exception as e:\n'
+            ' info["torch_error"]=str(e)\n'
+            'for p in ("/opt/rocm/.info/version","/opt/rocm/.info/version-dev"):\n'
+            ' f=pathlib.Path(p)\n'
+            ' if f.is_file():\n'
+            '  info["rocm_file"]=f.read_text().strip().splitlines()[0]\n'
+            '  break\n'
+            'for k in ("ROCM_VERSION","HIP_VERSION","TORCH_VERSION"):\n'
+            ' if os.environ.get(k): info[k.lower()]=os.environ[k]\n'
+            'def _dpkg_ver(*names):\n'
+            ' for n in names:\n'
+            '  try:\n'
+            '   r=subprocess.run(["dpkg-query","-W","-f=${Version}",n],'
+            'capture_output=True,text=True)\n'
+            '   if r.returncode==0 and r.stdout.strip():\n'
+            '    return r.stdout.strip()\n'
+            '  except Exception:\n'
+            '   pass\n'
+            ' return None\n'
+            'def _dpkg_scan(prefixes):\n'
+            ' try:\n'
+            '  r=subprocess.run(["dpkg-query","-W","-f=${Package}\\t${Version}\\n"],'
+            'capture_output=True,text=True)\n'
+            ' except Exception:\n'
+            '  return {}\n'
+            ' found={}\n'
+            ' for line in (r.stdout or "").splitlines():\n'
+            '  if "\\t" not in line: continue\n'
+            '  pkg,ver=line.split("\\t",1)\n'
+            '  for pref in prefixes:\n'
+            '   if pkg==pref or pkg.startswith(pref+"-"):\n'
+            '    found.setdefault(pref,ver)\n'
+            ' return found\n'
+            'pkgs=_dpkg_scan(("rccl","miopen"))\n'
+            'info["system.library.rccl"]=_dpkg_ver("rccl") or pkgs.get("rccl")\n'
+            'info["system.library.miopen"]=('
+            '_dpkg_ver("miopen-hip","miopen") or pkgs.get("miopen"))\n'
+            'print(json.dumps(info))\n')
+        for py in ('python3', 'python'):
+            result = cls._run_capture('docker', 'run', '--rm', '--network',
+                                      'none', '--entrypoint', py, base_image,
+                                      '-c', script)
+            if result.returncode == 0 and result.stdout.strip():
+                try:
+                    return json.loads(result.stdout.strip().splitlines()[-1])
+                except json.JSONDecodeError:
+                    continue
+        return {}
+
+    @classmethod
+    def _probe_via_history(cls, base_image: str) -> dict:
+        """Parse build ARGs from docker history (no container start)."""
+        result = cls._run_capture('docker', 'history', '--no-trunc',
+                                  '--format', '{{.CreatedBy}}', base_image)
+        if result.returncode != 0:
+            return {}
+        text = result.stdout
+        info = {}
+        for key, pattern in (
+            ('rocm', r'ROCM_VERSION=([0-9]+(?:\.[0-9]+)*)'),
+            ('python', r'PYTHON_VERSION=([0-9]+(?:\.[0-9]+)*)'),
+            ('ubuntu',
+             r'org\.opencontainers\.image\.version=([0-9]+(?:\.[0-9]+)*)'),
+        ):
+            matches = re.findall(pattern, text)
+            if matches:
+                # docker history lists newest layers first.
+                info[key] = matches[0]
+        return info
+
+    @classmethod
+    def _probe_via_create_cp(cls, base_image: str) -> dict:
+        """Copy version files out of a created (not started) container."""
+        import tempfile
+        create = cls._run_capture('docker', 'create', base_image)
+        if create.returncode != 0:
+            return {}
+        cid = create.stdout.strip()
+        info = {}
+        try:
+            with tempfile.TemporaryDirectory() as tmp:
+                dest = os.path.join(tmp, 'version')
+                for src in ('/opt/rocm/.info/version',
+                            '/opt/rocm/.info/version-dev'):
+                    result = cls._run_capture('docker', 'cp', f'{cid}:{src}',
+                                              dest)
+                    if result.returncode == 0 and os.path.isfile(dest):
+                        with open(dest, 'r', encoding='utf-8') as f:
+                            line = f.read().strip().splitlines()
+                        if line:
+                            info['rocm_file'] = line[0].strip()
+                            break
+        finally:
+            cls._run_capture('docker', 'rm', '-f', cid)
+        return info
+
+    @classmethod
+    def probe_base_image_versions(cls, base_image: str) -> dict:
+        """Discover rocm/python/torch without needing AMD GPU.
+
+        Methods (in order):
+        1. docker run --entrypoint python -c ...  (CPU-only, no --device)
+        2. docker history --no-trunc parse ROCM_VERSION/PYTHON_VERSION
+        3. docker create + docker cp /opt/rocm/.info/version
+        """
+        probed = {}
+        entry = cls._probe_via_entrypoint(base_image)
+        history = cls._probe_via_history(base_image)
+        copied = cls._probe_via_create_cp(base_image)
+        probed.update(history)
+        probed.update(copied)
+        probed.update(entry)
+
+        rocm = (
+            probed.get('rocm_file') or probed.get('rocm_version')
+            or probed.get('rocm') or probed.get('torch_hip')
+            or probed.get('hip_version'))
+        python_ver = probed.get('python')
+        torch_ver = probed.get('torch') or probed.get('torch_version')
+        ubuntu_ver = probed.get('ubuntu')
+        # Keep dpkg package versions as-is (may contain '~', e.g. 2.27.7.70201-81~22.04).
+        rccl_ver = probed.get('system.library.rccl')
+        miopen_ver = probed.get('system.library.miopen')
+
+        versions = {
+            'rocm': cls._normalize_version(rocm) if rocm else None,
+            'python':
+            cls._normalize_version(python_ver) if python_ver else None,
+            'torch': cls._normalize_version(torch_ver) if torch_ver else None,
+            'ubuntu':
+            cls._normalize_version(ubuntu_ver) if ubuntu_ver else None,
+            'system.library.rccl': rccl_ver or None,
+            'system.library.miopen': miopen_ver or None,
+        }
+        print('Probed AMD base image versions:')
+        for key, value in versions.items():
+            print(f'  {key}: {value or "unknown"}')
+        return versions
+
+    def generate_dockerfile(self) -> str:
+        with open('docker/Dockerfile.amd', 'r') as f:
+            content = f.read()
+        content = content.replace('{base_image}', self.args.base_image)
+        content = content.replace('{base_image_tag}', self.args.base_image_tag)
+        content = content.replace('{modelscope_branch}',
+                                  self.args.modelscope_branch)
+        content = content.replace('{cur_time}', formatted_time)
+        return content
+
+    def image(self) -> str:
+        ubuntu = getattr(self.args, 'amd_ubuntu_version',
+                         None) or self.args.ubuntu_version
+        rocm = getattr(self.args, 'amd_rocm_version', None)
+        py_tag = getattr(self.args, 'amd_python_tag', None) or getattr(
+            self.args, 'python_tag', None)
+        torch = getattr(self.args, 'amd_torch_version', None)
+        if not (rocm and py_tag and torch):
+            raise RuntimeError(
+                'AMD image tag requires probed rocm/python/torch versions. '
+                f'Got rocm={rocm}, python={py_tag}, torch={torch}')
+        return (f'{docker_registry}:ubuntu{ubuntu}-rocm{rocm}-{py_tag}-'
+                f'torch{torch}-{self.args.modelscope_version}-test')
+
+    def _log_base_image_info(self) -> int:
+        base_image = self.args.base_image
+        print('=' * 60)
+        print(f'AMD ROCm base image: {base_image}')
+        print(f'AMD ROCm base image tag: {self.args.base_image_tag}')
+        print('=' * 60)
+        ret = self.run_cmd('docker', 'pull', base_image)
+        if ret != 0:
+            return ret
+        result = self._run_capture(
+            'docker', 'image', 'inspect', base_image,
+            '--format={{.Id}} {{if index .RepoDigests 0}}'
+            '{{index .RepoDigests 0}}{{else}}local-only{{end}}')
+        if result.returncode == 0:
+            print(f'AMD base image resolved: {result.stdout.strip()}')
+        else:
+            print(f'AMD base image inspect warning: {result.stderr.strip()}')
+
+        versions = self.probe_base_image_versions(base_image)
+        if not versions.get('rocm') or not versions.get(
+                'python') or not versions.get('torch'):
+            print('ERROR: failed to probe rocm/python/torch from base image')
+            return 1
+        self.args.amd_rocm_version = versions['rocm']
+        self.args.amd_torch_version = versions['torch']
+        self.args.amd_python_tag = self._python_tag_from_version(
+            versions['python'])
+        if versions.get('ubuntu'):
+            self.args.amd_ubuntu_version = versions['ubuntu']
+        else:
+            self.args.amd_ubuntu_version = self.args.ubuntu_version
+        print(f'AMD output image tag will be: {self.image()}')
+        print('=' * 60)
+        return 0
+
+    def build(self) -> int:
+        ret = self._log_base_image_info()
+        if ret != 0:
+            return ret
+        return self.run_cmd('docker', 'build', '-t', self.image(), '-f',
+                            'Dockerfile', '.')
+
+    def push(self):
+        image_name = self.image()
+        ret = self.run_cmd('docker', 'push', image_name)
+        if ret != 0:
+            return ret
+        ubuntu = self.args.amd_ubuntu_version
+        rocm = self.args.amd_rocm_version
+        py_tag = self.args.amd_python_tag
+        torch = self.args.amd_torch_version
+        image_tag2 = (f'{docker_registry}:ubuntu{ubuntu}-rocm{rocm}-{py_tag}-'
+                      f'torch{torch}-{self.args.modelscope_version}-'
+                      f'{formatted_time}-test')
+        ret = self.run_cmd('docker', 'tag', image_name, image_tag2)
+        if ret != 0:
+            return ret
+        print(f'AMD image timestamp tag: {image_tag2}')
+        return self.run_cmd('docker', 'push', image_tag2)
+
+
 class AscendImageBuilder(StableGPUImageBuilder):

+    _DEFAULT_TORCH_VERSION = '2.9.0'
+    _DEFAULT_TORCHVISION_VERSION = '0.24.0'
+    _DEFAULT_TORCHAUDIO_VERSION = '2.9.0'
+    _DEFAULT_TORCH_NPU_VERSION = '2.9.0.post2'
+    _DEFAULT_VLLM_VERSION = '0.18.0'
+    _DEFAULT_VLLM_ASCEND_VERSION = '0.18.0'
+    _DEFAULT_TRITON_ASCEND_VERSIONS = {
+        '8.5': '3.2.0',
+        '9.0': '3.2.1',
+    }
    _CANN_VERSION_PATTERN = re.compile(r'^\d+(?:\.[0-9A-Za-z]+)+$')
    _OS_TAG_PATTERN = re.compile(r'^[A-Za-z]+[0-9][0-9A-Za-z.]*$')
+    _PYTHON_TAG_PATTERN = re.compile(r'^py\d+\.\d+$', re.IGNORECASE)
+    _TORCH_NPU_VERSION_PATTERN = re.compile(
+        r'^(?P<torch_version>\d+\.\d+\.\d+)(?:\.post\d+)?$')

    @staticmethod
    def _normalize_arch(arch: str = None) -> str:
        arch = arch or platform.machine()
        arch = arch.lower()
        arch_mapping = {
-            'x86': 'x86',
-            'x86_64': 'x86',
-            'amd64': 'x86',
-            'arm': 'arm',
-            'aarch64': 'arm',
-            'arm64': 'arm',
+            'x86': 'x86_64',
+            'x86_64': 'x86_64',
+            'amd64': 'x86_64',
+            'arm': 'aarch64',
+            'aarch64': 'aarch64',
+            'arm64': 'aarch64',
        }
        if arch not in arch_mapping:
            raise ValueError(f'Unsupported architecture: {arch}. '
@@ -536,34 +928,130 @@ class AscendImageBuilder(StableGPUImageBuilder):

        cann_version = parts[0]
        os_tag = parts[2]
+        python_tag = parts[3]
        if not cls._CANN_VERSION_PATTERN.fullmatch(cann_version):
            raise ValueError(f'Invalid CANN version in Ascend base image tag: '
                             f'{cann_version}')
        if not cls._OS_TAG_PATTERN.fullmatch(os_tag):
            raise ValueError(
                f'Invalid OS tag in Ascend base image tag: {os_tag}')
+        if not cls._PYTHON_TAG_PATTERN.fullmatch(python_tag):
+            raise ValueError(
+                f'Invalid Python tag in Ascend base image tag: {python_tag}')

-        return cann_version, f'CANN{cann_version}', os_tag
+        return cann_version, f'CANN{cann_version}', os_tag, python_tag
+
+    @staticmethod
+    def _get_os_family(os_tag: str) -> str:
+        os_tag = os_tag.lower()
+        if os_tag.startswith('ubuntu'):
+            return 'ubuntu'
+        if os_tag.startswith('openeuler'):
+            return 'openeuler'
+        raise ValueError(f'Unsupported Ascend base image OS tag: {os_tag}. '
+                         'Supported OS families are Ubuntu and openEuler.')
+
+    @classmethod
+    def _init_torch_versions(cls, args) -> None:
+        torch_version_specified = args.torch_version is not None
+        torchvision_version_specified = args.torchvision_version is not None
+        torchaudio_version_specified = args.torchaudio_version is not None
+
+        if torch_version_specified:
+            if (not torchvision_version_specified
+                    or not torchaudio_version_specified):
+                raise ValueError(
+                    'When overriding --torch_version for an Ascend image, also '
+                    'pass matching --torchvision_version and '
+                    '--torchaudio_version.')
+        elif torchvision_version_specified or torchaudio_version_specified:
+            raise ValueError(
+                '--torchvision_version and --torchaudio_version require an '
+                'explicit --torch_version for an Ascend image.')
+
+        args.torch_version = args.torch_version or cls._DEFAULT_TORCH_VERSION
+        args.torchvision_version = (
+            args.torchvision_version or cls._DEFAULT_TORCHVISION_VERSION)
+        args.torchaudio_version = (
+            args.torchaudio_version or cls._DEFAULT_TORCHAUDIO_VERSION)
+        args.torch_npu_version = (
+            args.torch_npu_version or cls._DEFAULT_TORCH_NPU_VERSION)
+
+        match = cls._TORCH_NPU_VERSION_PATTERN.fullmatch(
+            args.torch_npu_version)
+        if not match:
+            raise ValueError('Invalid --torch_npu_version. Expected '
+                             '<major>.<minor>.<patch> or '
+                             '<major>.<minor>.<patch>.post<revision>.')
+        if args.torch_version != match.group('torch_version'):
+            raise ValueError(
+                '--torch_version must exactly match the base version of '
+                f'--torch_npu_version, got torch={args.torch_version} and '
+                f'torch_npu={args.torch_npu_version}.')
+
+    @classmethod
+    def _init_component_versions(cls, args) -> None:
+        args.vllm_version = args.vllm_version or cls._DEFAULT_VLLM_VERSION
+        args.vllm_ascend_version = (
+            args.vllm_ascend_version or cls._DEFAULT_VLLM_ASCEND_VERSION)
+        args.vllm_git_ref = cls._get_vllm_git_ref(args.vllm_version)
+        args.vllm_ascend_git_ref = cls._get_vllm_git_ref(
+            args.vllm_ascend_version)
+
+        if not args.triton_ascend_version:
+            cann_series = '.'.join(args.cann_version.split('.')[:2])
+            try:
+                args.triton_ascend_version = (
+                    cls._DEFAULT_TRITON_ASCEND_VERSIONS[cann_series])
+            except KeyError as e:
+                raise ValueError('No default triton-ascend version for CANN '
+                                 f'{args.cann_version}. Please pass '
+                                 '--triton_ascend_version explicitly.') from e
+
+    @staticmethod
+    def _get_vllm_git_ref(version: str) -> str:
+        return version if version.startswith('v') else f'v{version}'

    def init_args(self, args) -> Any:
        if not args.base_image:
            # Reuse the prebuilt vllm-ascend image to avoid rebuilding its stack.
            args.base_image = 'quay.io/ascend/cann:8.5.1-a3-ubuntu22.04-py3.11'
+        self._init_torch_versions(args)
        args.arch = self._normalize_arch(args.arch)
        args.atlas_hardware = self._get_atlas_hardware(args.soc_version)
-        args.cann_version, args.cann_version_tag, args.os_tag = (
-            self._get_cann_os_tags(args.base_image))
+        (args.cann_version, args.cann_version_tag, args.os_tag,
+         args.ascend_python_tag) = (
+             self._get_cann_os_tags(args.base_image))
+        self._get_os_family(args.os_tag)
+        self._init_component_versions(args)
        return super().init_args(args)

+    def _generate_python_tag(self, _python_version: str) -> str:
+        return self.args.ascend_python_tag
+
    def generate_dockerfile(self) -> str:
        extra_content = """
-RUN pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
+RUN export PIP_EXTRA_INDEX_URL=https://pypi.org/simple && \
+    pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
 """
        with open('docker/Dockerfile.ascend', 'r') as f:
            content = f.read()
            content = content.replace('{base_image}', self.args.base_image)
            content = content.replace('{soc_version}', self.args.soc_version)
            content = content.replace('{cann_version}', self.args.cann_version)
+            content = content.replace('{torch_version}',
+                                      self.args.torch_version)
+            content = content.replace('{torchvision_version}',
+                                      self.args.torchvision_version)
+            content = content.replace('{torchaudio_version}',
+                                      self.args.torchaudio_version)
+            content = content.replace('{torch_npu_version}',
+                                      self.args.torch_npu_version)
+            content = content.replace('{vllm_git_ref}', self.args.vllm_git_ref)
+            content = content.replace('{vllm_ascend_git_ref}',
+                                      self.args.vllm_ascend_git_ref)
+            content = content.replace('{triton_ascend_version}',
+                                      self.args.triton_ascend_version)
            content = content.replace('{extra_content}', extra_content)
            content = content.replace('{cur_time}', formatted_time)
            content = content.replace('{install_ms_deps}', 'False')
@@ -577,11 +1065,12 @@ RUN pip install --no-cache-dir -U icecream soundfile pybind11 py-spy
        return content

    def image(self) -> str:
-        return (
-            f'{docker_registry}:{self.args.swift_branch}-'
-            f'{self.args.atlas_hardware}-{self.args.python_tag}-'
-            f'{self.args.cann_version_tag}-{self.args.os_tag}-{self.args.arch}'
-        )
+        tag = (f'{self.args.swift_branch}-{self.args.cann_version_tag}-'
+               f'torch_npu{self.args.torch_npu_version}-'
+               f'{self.args.atlas_hardware}-{self.args.os_tag}-'
+               f'{self.args.python_tag}-'
+               f'{self.args.arch}')
+        return f'{docker_registry}:{tag.lower()}'

    def push(self):
        return 0
@@ -593,6 +1082,7 @@ parser.add_argument('--image_type', type=str)
 parser.add_argument('--python_version', type=str, default='3.12.13')
 parser.add_argument('--ubuntu_version', type=str, default='22.04')
 parser.add_argument('--torch_version', type=str, default=None)
+parser.add_argument('--torch_npu_version', type=str, default=None)
 parser.add_argument('--torchvision_version', type=str, default=None)
 parser.add_argument('--cuda_version', type=str, default=None)
 parser.add_argument('--ci_image', type=int, default=0)
@@ -600,6 +1090,8 @@ parser.add_argument('--torchaudio_version', type=str, default=None)
 parser.add_argument('--optimum_version', type=str, default=None)
 parser.add_argument('--tf_version', type=str, default=None)
 parser.add_argument('--vllm_version', type=str, default=None)
+parser.add_argument('--vllm_ascend_version', type=str, default=None)
+parser.add_argument('--triton_ascend_version', type=str, default=None)
 parser.add_argument('--lmdeploy_version', type=str, default=None)
 parser.add_argument('--flashattn_version', type=str, default=None)
 parser.add_argument('--autogptq_version', type=str, default=None)
@@ -610,6 +1102,12 @@ parser.add_argument('--megatron_branch', type=str, default='v0.15.3')
 parser.add_argument('--mindspeed_branch', type=str, default='core_r0.15.3')
 parser.add_argument('--soc_version', type=str, default='ascend910_9391')
 parser.add_argument('--arch', type=str, choices=['x86', 'arm'], default=None)
+parser.add_argument(
+    '--base_image_tag',
+    type=str,
+    default=None,
+    help='Optional AMD ROCm override tag. Default: auto-resolve newest '
+    'concrete vllm/vllm-openai-rocm release tag from Docker Hub.')
 parser.add_argument('--dry_run', type=int, default=0)
 args = parser.parse_args()

@@ -621,6 +1119,8 @@ elif args.image_type.lower() == 'stable':
    builder_cls = [StableCPUImageBuilder, StableGPUImageBuilder]
 elif args.image_type.lower() == 'ascend':
    builder_cls = [AscendImageBuilder]
+elif args.image_type.lower() == 'amd':
+    builder_cls = [AmdImageBuilder]
 elif args.image_type.lower() == 'latest':
    builder_cls = [LatestGPUImageBuilder]
 else:
--- a/docker/scripts/modelscope_env_init.sh
+++ b/docker/scripts/modelscope_env_init.sh
@@ -47,5 +47,4 @@ else
 fi

 pip config set global.index-url https://mirrors.cloud.aliyuncs.com/pypi/simple
-pip config set global.extra-index-url https://pypi.org/simple
 pip config set install.trusted-host mirrors.cloud.aliyuncs.com