diff --git a/.ci/docker/README.md b/.ci/docker/README.md
new file mode 100644
index 0000000000..7771f015f9
--- /dev/null
+++ b/.ci/docker/README.md
@@ -0,0 +1,91 @@
+# torch-npu CI Docker Images
+
+本目录管理 torch-npu 项目的 CI Docker 镜像，包括**构建镜像 (builder)** 和**测试镜像 (test)** 两类，每类分别支持 x86_64 和 aarch64 架构。
+
+## 镜像类型
+
+| 类型 | 基座 | 用途 |
+|------|------|------|
+| **builder** | manylinux2_28-builder | 编译构建 torch-npu wheel 包，包含完整编译工具链 |
+| **test** | ubuntu:22.04 | CI 单元测试运行环境，包含 PyTorch CPU、CANN runtime、triton-ascend 和测试框架 |
+
+## 目录结构
+
+```
+.ci/docker/
+├── README.md
+├── requirements-builder.txt      # Builder 镜像 pip 依赖
+├── requirements-test.txt         # Test 镜像 pip 依赖
+├── docker_build.sh               # 构建入口脚本
+├── common/                       # 共享安装脚本
+│   ├── install_cann.sh           # 安装 CANN toolkit (支持 A1/A2/A3)
+│   ├── install_triton.sh         # 安装 triton-ascend (需传 Python 版本)
+│   ├── install_obs.sh            # 安装华为 OBS util
+├── builder/
+│   ├── Dockerfile.x86_64
+│   └── Dockerfile.aarch64
+└── test/
+    ├── Dockerfile.x86_64
+    └── Dockerfile.aarch64
+```
+
+## 快速构建
+
+```bash
+# Builder 镜像 (不含 CANN)
+./docker_build.sh torch-npu-builder-x86_64-torch2.7.1
+./docker_build.sh torch-npu-builder-aarch64-torch2.7.1
+
+# Test 镜像 (含 CANN)
+./docker_build.sh torch-npu-test-x86_64-cann-a1-py3.10-torch2.7.1
+./docker_build.sh torch-npu-test-aarch64-cann-a2-py3.10-torch2.7.1
+```
+
+## Tag 命名规范
+
+参考上游 PyTorch `pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11` 模式，tag 即为最终镜像名：
+
+**Builder**（不含 CANN）：
+```
+torch-npu-builder-<ARCH>-torch<PYTORCH_VERSION>
+```
+```
+./docker_build.sh torch-npu-builder-x86_64-torch2.7.1
+#                   ^          ^       ^     ^
+#                   |          |       |     └── PyTorch 版本 (torch2.7.1)
+#                   |          |       └── 架构
+#                   |          └── 镜像类型
+#                   └── 固定前缀
+```
+
+**Test**（含 CANN runtime）：
+```
+torch-npu-test-<ARCH>-cann<CHIP>-py<PYTHON_VERSION>-torch<PYTORCH_VERSION>
+```
+```
+./docker_build.sh torch-npu-test-x86_64-cann-a1-py3.10-torch2.7.1
+#                   ^         ^       ^    ^  ^     ^     ^
+#                   |         |       |    |  |     |     └── PyTorch 版本
+#                   |         |       |    |  |     └── torch 前缀
+#                   |         |       |    |  └── Python 版本
+#                   |         |       |    └── py 前缀
+#                   |         |       └── CANN 芯片 (A1/A2/A3)
+#                   |         └── 架构
+#                   └── 镜像类型
+```
+
+| 字段 | 可选值 |
+|------|--------|
+| IMAGE_TYPE | builder, test |
+| ARCH | x86_64, aarch64 |
+| CHIP | A1 (Ascend 910), A2 (Ascend 910b), A3 (仅 test) |
+| PYTHON_VERSION | 3.10 (仅 test) |
+| PYTORCH_VERSION | 2.7.1 |
+
+## CANN 芯片映射
+
+| CANN_CHIP | 芯片 | CANN 版本 |
+|-----------|------|----------|
+| A1 | Ascend 910 | 9.1.0 |
+| A2 | Ascend 910b | 8.5.0 (x86_64) / 9.1.0 (aarch64) |
+| A3 | Ascend A3 | 9.0.0-beta.1 (x86_64) / 9.0.0-beta.2 (aarch64) |
diff --git a/.ci/docker/builder/Dockerfile.aarch64 b/.ci/docker/builder/Dockerfile.aarch64
new file mode 100644
index 0000000000..7cc382123c
--- /dev/null
+++ b/.ci/docker/builder/Dockerfile.aarch64
@@ -0,0 +1,71 @@
+FROM pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-2.7
+
+ARG PYTORCH_VERSION=2.7.1
+
+ENV PATH=/usr/local/bin:$PATH
+ENV AUDITWHEEL_PLAT=manylinux_2_28_aarch64
+ENV ETCD_UNSUPPORTED_ARCH=arm64
+ENV PYTORCH_VERSION=${PYTORCH_VERSION}
+
+COPY requirements-builder.txt /opt/buildtools/
+
+# Set pip & python symlinks
+RUN cd /usr/local/bin \
+    && ln -sf /opt/_internal/cpython-3.9.21/bin/pip3.9 pip3.9 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/pip3.10 pip3.10 \
+    && ln -sf /opt/_internal/cpython-3.11.11/bin/pip3.11 pip3.11 \
+    && ln -sf /opt/_internal/cpython-3.12.9/bin/pip3.12 pip3.12 \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/pip3.13 pip3.13 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/pip3.10 pip3 \
+    && ln -sf /opt/_internal/cpython-3.9.21/bin/python3.9 python3.9 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/python3.10 python3.10 \
+    && ln -sf /opt/_internal/cpython-3.11.11/bin/python3.11 python3.11 \
+    && ln -sf /opt/_internal/cpython-3.12.9/bin/python3.12 python3.12 \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/python3.13 python3.13 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/python3.10 python3
+
+# Install PyTorch from official source, then build requirements from PyPI for each python version
+RUN pip3.9 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+        && pip3.9 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.10 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+        && pip3.10 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.11 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+        && pip3.11 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.12 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+        && pip3.12 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.13 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+        && pip3.13 install --no-cache-dir auditwheel==5.4.0 \
+        && pip3.13 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/auditwheel /usr/local/bin/auditwheel
+
+# Install system build tools
+RUN echo "alias ll='ls -l --color=auto'" >> /root/.bashrc \
+    && yum install -y vim-common --disablerepo=ius \
+    && yum install -y ninja-build binutils lld mold dos2unix gcc gcc-c++ make cmake3 wget tar unzip elfutils java-1.8.0-openjdk-devel \
+    && cd /tmp \
+    && wget https://github.com/ccache/ccache/releases/download/v4.10/ccache-4.10.tar.gz \
+    && tar -xzf ccache-4.10.tar.gz \
+    && cd ccache-4.10 \
+    && mkdir build \
+    && cd build \
+    && cmake3 .. \
+    && make -j$(nproc) \
+    && make install \
+    && cd /tmp \
+    && rm -rf ccache-4.10* \
+    && ccache --version \
+    && wget https://github.com/etcd-io/etcd/releases/download/v3.4.3/etcd-v3.4.3-linux-arm64.tar.gz \
+    && tar -zxf etcd-v3.4.3-linux-arm64.tar.gz \
+    && mv etcd-v3.4.3-linux-arm64/etcd /usr/local/bin/ \
+    && pip3.10 install python-etcd \
+    && etcd --version \
+    && yum update -y \
+    && yum clean all
+
+# Set timezone
+RUN rm -f /etc/localtime \
+    && cp /usr/share/zoneinfo/Asia/Shanghai /etc/localtime \
+    && echo 'Asia/Shanghai' >/etc/timezone \
+    && echo "export TZ='Asia/Shanghai'" >>/etc/profile
+
+WORKDIR /home
diff --git a/.ci/docker/builder/Dockerfile.x86_64 b/.ci/docker/builder/Dockerfile.x86_64
new file mode 100644
index 0000000000..551adc2626
--- /dev/null
+++ b/.ci/docker/builder/Dockerfile.x86_64
@@ -0,0 +1,82 @@
+FROM pytorch/manylinux2_28-builder:cpu-2.7
+
+ARG PYTORCH_VERSION=2.7.1
+
+ENV PATH=/usr/local/bin:$PATH
+ENV AUDITWHEEL_PLAT=manylinux_2_28_x86_64
+ENV PYTORCH_VERSION=${PYTORCH_VERSION}
+
+COPY requirements-builder.txt /opt/buildtools/
+
+# Set pip & python symlinks
+RUN cd /usr/local/bin \
+    && ln -sf /opt/_internal/cpython-3.9.21/bin/pip3.9 pip3.9 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/pip3.10 pip3.10 \
+    && ln -sf /opt/_internal/cpython-3.11.11/bin/pip3.11 pip3.11 \
+    && ln -sf /opt/_internal/cpython-3.12.9/bin/pip3.12 pip3.12 \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/pip3.13 pip3.13 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/pip3.10 pip3 \
+    && ln -sf /opt/_internal/cpython-3.9.21/bin/python3.9 python3.9 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/python3.10 python3.10 \
+    && ln -sf /opt/_internal/cpython-3.11.11/bin/python3.11 python3.11 \
+    && ln -sf /opt/_internal/cpython-3.12.9/bin/python3.12 python3.12 \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/python3.13 python3.13 \
+    && ln -sf /opt/_internal/cpython-3.10.16/bin/python3.10 python3
+
+# Install PyTorch from official source, then build requirements from PyPI for each python version
+RUN pip3.9 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION}+cpu \
+        && pip3.9 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.10 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION}+cpu \
+        && pip3.10 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.11 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION}+cpu \
+        && pip3.11 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.12 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION}+cpu \
+        && pip3.12 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && pip3.13 install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION}+cpu \
+        && pip3.13 install --no-cache-dir auditwheel==5.4.0 \
+        && pip3.13 install --no-cache-dir -r /opt/buildtools/requirements-builder.txt \
+    && ln -sf /opt/_internal/cpython-3.13.2/bin/auditwheel /usr/local/bin/auditwheel
+
+# Install system build tools
+RUN yum remove -y ius-release epel-release 2>/dev/null || true \
+    && rm -rf /etc/yum.repos.d/ius*.repo /etc/yum.repos.d/epel*.repo \
+    && yum clean all && rm -rf /var/cache/dnf /var/cache/yum \
+    && echo "alias ll='ls -l --color=auto'" >> /root/.bashrc \
+    && yum install -y vim-common --disablerepo=ius \
+    && yum install -y binutils lld dos2unix gcc gcc-c++ make cmake3 wget tar unzip elfutils java-1.8.0-openjdk-devel \
+    && cd /tmp \
+    && wget -q https://github.com/ninja-build/ninja/releases/download/v1.12.1/ninja-linux.zip \
+    && unzip ninja-linux.zip \
+    && cp ninja /usr/local/bin/ && chmod +x /usr/local/bin/ninja \
+    && cd /tmp \
+    && wget -q https://github.com/rui314/mold/archive/refs/tags/v2.32.1.tar.gz \
+    && tar -xf v2.32.1.tar.gz \
+    && cd mold-2.32.1 \
+    && cmake -DCMAKE_BUILD_TYPE=Release -DMOLD_MOSTLY_STATIC=ON . \
+    && make -j$(nproc) && make install \
+    && cd /tmp \
+    && wget https://github.com/ccache/ccache/releases/download/v4.10/ccache-4.10.tar.gz \
+    && tar -xzf ccache-4.10.tar.gz \
+    && cd ccache-4.10 \
+    && mkdir build \
+    && cd build \
+    && cmake3 .. \
+    && make -j$(nproc) \
+    && make install \
+    && cd /tmp && rm -rf /tmp/* \
+    && ninja --version && mold --version && ccache --version \
+    && wget https://github.com/etcd-io/etcd/releases/download/v3.4.3/etcd-v3.4.3-linux-amd64.tar.gz \
+    && tar -zxf etcd-v3.4.3-linux-amd64.tar.gz \
+    && mv etcd-v3.4.3-linux-amd64/etcd /usr/local/bin/ \
+    && pip3.10 install python-etcd \
+    && etcd --version \
+    && yum clean all \
+    && yum update -y
+
+# Set timezone
+RUN rm -f /etc/localtime \
+    && cp /usr/share/zoneinfo/Asia/Shanghai /etc/localtime \
+    && echo 'Asia/Shanghai' >/etc/timezone \
+    && echo "export TZ='Asia/Shanghai'" >>/etc/profile
+
+WORKDIR /home
diff --git a/.ci/docker/common/install_cann.sh b/.ci/docker/common/install_cann.sh
new file mode 100755
index 0000000000..4af49bbb89
--- /dev/null
+++ b/.ci/docker/common/install_cann.sh
@@ -0,0 +1,109 @@
+#!/usr/bin/bash
+# Install CANN toolkit for Ascend NPU.
+# Usage: CANN_CHIP=A1 ./install_cann.sh
+#   CANN_CHIP: A1 (Ascend 910), A2 (Ascend 910b), A3 (Ascend A3)
+# Automatically detects architecture (x86_64 / aarch64).
+
+set -e
+
+CANN_CHIP="${CANN_CHIP:-A1}"
+ARCH=$(uname -m)
+
+BASE_URL="https://pytorch-package.obs.cn-north-4.myhuaweicloud.com/pta/cann-package"
+CANN_BASE_URL="https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%209.1.T1"
+
+case "${ARCH}_${CANN_CHIP}" in
+  # x86_64
+  x86_64_A1)
+    TOOLKIT_URL="${BASE_URL}/20260513/Ascend-cann-toolkit_9.1.0_linux-x86_64.run"
+    OPS_URL="${BASE_URL}/20260513/Ascend-cann-910-ops_9.1.0_linux-x86_64.run"
+    NNAL_URL="${BASE_URL}/20260513/Ascend-cann-nnal_9.1.0_linux-x86_64.run"
+    OPS_GLOB="Ascend-cann-910*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  x86_64_A2)
+    TOOLKIT_URL="${CANN_BASE_URL}/Ascend-cann-toolkit_9.1.0-beta.1_linux-x86_64.run"
+    OPS_URL="${CANN_BASE_URL}/Ascend-cann-910b-ops_9.1.0-beta.1_linux-x86_64.run"
+    NNAL_URL="${CANN_BASE_URL}/Ascend-cann-nnal_9.1.0-beta.1_linux-x86_64.run"
+    OPS_GLOB="Ascend-cann-910b-ops*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  x86_64_A3)
+    TOOLKIT_URL="${CANN_BASE_URL}/Ascend-cann-toolkit_9.1.0-beta.1_linux-x86_64.run"
+    OPS_URL="${CANN_BASE_URL}/Ascend-cann-A3-ops_9.1.0-beta.1_linux-x86_64.run"
+    NNAL_URL="${CANN_BASE_URL}/Ascend-cann-nnal_9.1.0-beta.1_linux-x86_64.run"
+    OPS_GLOB="Ascend-cann-A3-ops*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  # aarch64
+  aarch64_A1)
+    TOOLKIT_URL="${BASE_URL}/20260302/Ascend-cann-toolkit_9.0.0-beta.1_linux-aarch64.run"
+    OPS_URL="${BASE_URL}/20260302/Ascend-cann-910b-ops_9.0.0-beta.1_linux-aarch64.run"
+    NNAL_URL="${BASE_URL}/20260302/Ascend-cann-nnal_9.0.0-beta.1_linux-aarch64.run"
+    OPS_GLOB="Ascend-cann-910b*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  aarch64_A2)
+    TOOLKIT_URL="${CANN_BASE_URL}/Ascend-cann-toolkit_9.1.0-beta.1_linux-aarch64.run"
+    OPS_URL="${CANN_BASE_URL}/Ascend-cann-910b-ops_9.1.0-beta.1_linux-aarch64.run"
+    NNAL_URL="${CANN_BASE_URL}/Ascend-cann-nnal_9.1.0-beta.1_linux-aarch64.run"
+    OPS_GLOB="Ascend-cann-910b-ops*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  aarch64_A3)
+    TOOLKIT_URL="${CANN_BASE_URL}/Ascend-cann-toolkit_9.1.0-beta.1_linux-aarch64.run"
+    OPS_URL="${CANN_BASE_URL}/Ascend-cann-A3-ops_9.1.0-beta.1_linux-aarch64.run"
+    NNAL_URL="${CANN_BASE_URL}/Ascend-cann-nnal_9.1.0-beta.1_linux-aarch64.run"
+    OPS_GLOB="Ascend-cann-A3-ops*"
+    SET_ENV_PATH="/usr/local/Ascend/cann/set_env.sh"
+    ;;
+  *)
+    echo "Unsupported combination: ${ARCH} + ${CANN_CHIP}"
+    exit 1
+    ;;
+esac
+
+echo "Installing CANN ${CANN_CHIP} for ${ARCH}..."
+
+echo "=== Creating HwHiAiUser user and group ==="
+groupadd -f HwHiAiUser
+id -u HwHiAiUser >/dev/null 2>&1 || useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash
+
+rm -rf cann
+mkdir -p cann && cd cann
+
+echo "=== Downloading CANN packages ==="
+curl -O "${TOOLKIT_URL}"
+curl -O "${OPS_URL}"
+curl -O "${NNAL_URL}"
+echo "Download complete."
+
+chmod +x Ascend-cann*.run
+
+echo "=== Installing CANN toolkit ==="
+./Ascend-cann-toolkit*.run --full --quiet --install-path=/usr/local/Ascend
+source "${SET_ENV_PATH}"
+echo "toolkit install success"
+
+echo "=== Installing CANN ops ==="
+./${OPS_GLOB}.run --install --quiet --install-path=/usr/local/Ascend
+echo "ops install success"
+
+echo "=== Installing CANN nnal ==="
+./Ascend-cann-nnal*.run --install --quiet --install-path=/usr/local/Ascend
+source /usr/local/Ascend/nnal/atb/set_env.sh
+echo "nnal install success"
+
+# Some CANN versions install to versioned paths (e.g. cann-9.0.0-beta.2)
+# instead of /usr/local/Ascend/cann/. Fix broken symlinks so runtime
+# sourcing of set_env.sh works.
+if [ ! -f /usr/local/Ascend/cann/set_env.sh ]; then
+  CANN_REAL_DIR=$(ls -d /usr/local/Ascend/cann-* 2>/dev/null | head -1)
+  if [ -n "${CANN_REAL_DIR}" ]; then
+    ln -sf "${CANN_REAL_DIR}" /usr/local/Ascend/cann
+    echo "Fixed: linked ${CANN_REAL_DIR} -> /usr/local/Ascend/cann"
+  fi
+fi
+
+rm -rf *
+echo "CANN ${CANN_CHIP} installation complete."
\ No newline at end of file
diff --git a/.ci/docker/common/install_obs.sh b/.ci/docker/common/install_obs.sh
new file mode 100755
index 0000000000..1acea84f2b
--- /dev/null
+++ b/.ci/docker/common/install_obs.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/bash
+# Install Huawei OBS util for object storage access.
+
+set -e
+
+ARCH=$(uname -m)
+case "${ARCH}" in
+  x86_64)  OBS_ARCH="amd64" ;;
+  aarch64) OBS_ARCH="arm64" ;;
+  *)       echo "Unsupported architecture: ${ARCH}"; exit 1 ;;
+esac
+
+OBS_URL="https://obs-community.obs.cn-north-1.myhuaweicloud.com/obsutil/current/obsutil_linux_${OBS_ARCH}.tar.gz"
+
+wget -q "${OBS_URL}"
+mkdir -p /usr/local/obsutil
+tar -zxf "obsutil_linux_${OBS_ARCH}.tar.gz" -C /usr/local/obsutil/
+rm -f "obsutil_linux_${OBS_ARCH}.tar.gz"
+ln -sf /usr/local/obsutil/obsutil_linux_${OBS_ARCH}_*/obsutil /usr/local/bin/obsutil
+
+echo "OBS util installed."
diff --git a/.ci/docker/common/install_triton.sh b/.ci/docker/common/install_triton.sh
new file mode 100755
index 0000000000..ed76bca16d
--- /dev/null
+++ b/.ci/docker/common/install_triton.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/bash
+# Install triton-ascend for NPU.
+# Usage: ./install_triton.sh <PYTHON_VERSION>
+#   PYTHON_VERSION: e.g. 3.10, 3.11, 3.12, 3.13
+
+set -e
+
+TRITON_VERSION="${TRITON_VERSION:-3.2.1}"
+PYTHON_VERSION="${1:?Usage: $0 <PYTHON_VERSION> (e.g. 3.10)}"
+
+ARCH=$(uname -m)
+PY_SHORT=$(echo "${PYTHON_VERSION}" | tr -d '.')
+
+TRITON_WHL="triton_ascend-${TRITON_VERSION}-cp${PY_SHORT}-cp${PY_SHORT}-manylinux_2_27_${ARCH}.manylinux_2_28_${ARCH}.whl"
+TRITON_URL="https://gitcode.com/Ascend/triton-ascend/releases/download/v${TRITON_VERSION}/${TRITON_WHL}"
+
+echo "Installing triton-ascend ${TRITON_VERSION} for Python ${PYTHON_VERSION} (${ARCH})..."
+pip3 install --no-cache-dir "${TRITON_URL}"
+echo "triton-ascend installed."
diff --git a/.ci/docker/docker_build.sh b/.ci/docker/docker_build.sh
new file mode 100755
index 0000000000..2183f141d6
--- /dev/null
+++ b/.ci/docker/docker_build.sh
@@ -0,0 +1,115 @@
+#!/usr/bin/bash
+# Build torch-npu CI Docker images.
+#
+# Usage:
+#   ./docker_build.sh <TAG>
+#
+# Builder: torch-npu-builder-<ARCH>-torch<PYTORCH_VERSION>
+# Test:    torch-npu-test-<ARCH>-cann<CHIP>-py<PYTHON_VERSION>-torch<PYTORCH_VERSION>
+#
+# Examples:
+#   ./docker_build.sh torch-npu-builder-x86_64-torch2.7.1
+#   ./docker_build.sh torch-npu-test-aarch64-cann-a2-py3.10-torch2.7.1
+#
+# Reference: pytorch/pytorch .ci/docker/build.sh
+
+set -ex
+
+tag="${1:?Usage: $0 <TAG>}"
+shift
+
+case "$tag" in
+  torch-npu-builder-x86_64-torch2.7.1)
+    IMAGE_TYPE=builder
+    ARCH=x86_64
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-builder-aarch64-torch2.7.1)
+    IMAGE_TYPE=builder
+    ARCH=aarch64
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-x86_64-cann-a1-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=x86_64
+    CANN_CHIP=A1
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-x86_64-cann-a2-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=x86_64
+    CANN_CHIP=A2
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-x86_64-cann-a3-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=x86_64
+    CANN_CHIP=A3
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-aarch64-cann-a1-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=aarch64
+    CANN_CHIP=A1
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-aarch64-cann-a2-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=aarch64
+    CANN_CHIP=A2
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  torch-npu-test-aarch64-cann-a3-py3.10-torch2.7.1)
+    IMAGE_TYPE=test
+    ARCH=aarch64
+    CANN_CHIP=A3
+    PYTHON_VERSION=3.10
+    PYTORCH_VERSION=2.7.1
+    ;;
+  *)
+    echo "Unknown tag: ${tag}"
+    echo "  Builder: torch-npu-builder-<x86_64|aarch64>-torch2.7.1"
+    echo "  Test:    torch-npu-test-<x86_64|aarch64>-cann<A1|A2|A3>-py3.10-torch2.7.1"
+    exit 1
+    ;;
+esac
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+DOCKERFILE="${SCRIPT_DIR}/${IMAGE_TYPE}/Dockerfile.${ARCH}"
+
+if [[ ! -f "${DOCKERFILE}" ]]; then
+  echo "Dockerfile not found: ${DOCKERFILE}"
+  exit 1
+fi
+
+BUILD_ARGS=(
+  --build-arg PYTORCH_VERSION="${PYTORCH_VERSION}"
+)
+if [[ -n "${CANN_CHIP:-}" ]]; then
+  BUILD_ARGS+=(--build-arg CANN_CHIP="${CANN_CHIP}")
+fi
+if [[ -n "${PYTHON_VERSION:-}" ]]; then
+  BUILD_ARGS+=(--build-arg PYTHON_VERSION="${PYTHON_VERSION}")
+fi
+
+TIMESTAMP="${TIMESTAMP:-$(date -u +%Y%m%d%H%M)}"
+IMAGE_TAG="${tag}-${TIMESTAMP}"
+
+echo "Building ${IMAGE_TAG} ..."
+echo "  Dockerfile: ${DOCKERFILE}"
+echo "  PyTorch:    ${PYTORCH_VERSION}"
+[[ -n "${PYTHON_VERSION:-}" ]] && echo "  Python:     ${PYTHON_VERSION}"
+[[ -n "${CANN_CHIP:-}" ]] && echo "  CANN chip:  ${CANN_CHIP}"
+
+docker build \
+  -f "${DOCKERFILE}" \
+  -t "${IMAGE_TAG}" \
+  "${BUILD_ARGS[@]}" \
+  "${SCRIPT_DIR}"
+
+echo "Image built: ${IMAGE_TAG}"
diff --git a/.ci/docker/requirements-builder.txt b/.ci/docker/requirements-builder.txt
new file mode 100644
index 0000000000..871634e0df
--- /dev/null
+++ b/.ci/docker/requirements-builder.txt
@@ -0,0 +1,5 @@
+numpy==1.26.4
+pybind11==2.13.1
+pyyaml
+setuptools==75.3.2
+wheel
diff --git a/.ci/docker/requirements-ci.txt b/.ci/docker/requirements-ci.txt
deleted file mode 100644
index 8602d4d0fa..0000000000
--- a/.ci/docker/requirements-ci.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-# Python dependencies required for unit tests
-
-mypy==1.9.0
-# Pin MyPy version because new errors are likely to appear with each release
-#Description: linter
-#Pinned versions: 1.9.0
-#test that import: test_typing.py, test_type_hints.py
diff --git a/.ci/docker/requirements-test.txt b/.ci/docker/requirements-test.txt
new file mode 100644
index 0000000000..ecb003b272
--- /dev/null
+++ b/.ci/docker/requirements-test.txt
@@ -0,0 +1,42 @@
+# Python dependencies required for CI unit tests
+
+-f https://data.pyg.org/whl/torch-2.7.1+cpu.html
+
+# Test frameworks
+pytest==8.1.1
+pytest-xdist
+pytest-subtests
+coverage
+hypothesis
+parameterized==0.9.0
+expecttest==0.1.3
+unittest-xml-reporting
+
+mypy==1.14.0
+
+ml-dtypes==0.5.1
+numpy==1.26.4
+pytest-timeout==2.3.1
+
+onnx==1.17.0
+onnxruntime==1.18.1
+onnxscript==0.2.2
+
+Pillow==10.3.0
+protobuf==3.20.2
+requests==2.32.0
+
+torch_geometric==2.5.3
+torch-scatter==2.1.2
+torchvision==0.22.1
+transformers==4.40.0
+
+tabulate==0.9.0
+importlib_metadata
+optree
+packaging
+psutil
+scipy
+z3-solver==4.13.0.0
+zstandard==0.25.0
+pulp==3.3.1
diff --git a/.ci/docker/test/Dockerfile.aarch64 b/.ci/docker/test/Dockerfile.aarch64
new file mode 100644
index 0000000000..1b44929e00
--- /dev/null
+++ b/.ci/docker/test/Dockerfile.aarch64
@@ -0,0 +1,54 @@
+FROM ubuntu:22.04
+
+ARG PYTORCH_VERSION=2.7.1
+ARG CANN_CHIP=A2
+ARG PYTHON_VERSION=3.10
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV TZ=Asia/Shanghai
+ENV PATH=/usr/local/bin:$PATH
+ENV PYTORCH_VERSION=${PYTORCH_VERSION}
+ENV CANN_CHIP=${CANN_CHIP}
+
+COPY common/ /opt/buildtools/
+COPY requirements-test.txt /opt/buildtools/
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        curl \
+        dos2unix \
+        gcc \
+        g++ \
+        git \
+        make \
+        python3 \
+        python3-dev \
+        python3-pip \
+        python3-venv \
+        tar \
+        tzdata \
+        unzip \
+        vim \
+        wget \
+    && ln -sf /usr/bin/python3 /usr/bin/python \
+    && rm -rf /var/lib/apt/lists/*
+
+# Upgrade pip/setuptools/wheel
+RUN python3 -m pip install --no-cache-dir --upgrade pip setuptools wheel
+
+# Install CANN and OBS
+RUN chmod -R 755 /opt/buildtools/* \
+    && dos2unix /opt/buildtools/* \
+    && /opt/buildtools/install_cann.sh \
+    && /opt/buildtools/install_obs.sh
+
+# Install triton-ascend
+RUN /opt/buildtools/install_triton.sh 3.10
+
+# Install PyTorch from official source, then test requirements from PyPI
+RUN python3 -m pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+    && python3 -m pip install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu -r /opt/buildtools/requirements-test.txt
+
+WORKDIR /home
diff --git a/.ci/docker/test/Dockerfile.x86_64 b/.ci/docker/test/Dockerfile.x86_64
new file mode 100644
index 0000000000..e698f04d6c
--- /dev/null
+++ b/.ci/docker/test/Dockerfile.x86_64
@@ -0,0 +1,54 @@
+FROM ubuntu:22.04
+
+ARG PYTORCH_VERSION=2.7.1
+ARG CANN_CHIP=A1
+ARG PYTHON_VERSION=3.10
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV TZ=Asia/Shanghai
+ENV PATH=/usr/local/bin:$PATH
+ENV PYTORCH_VERSION=${PYTORCH_VERSION}
+ENV CANN_CHIP=${CANN_CHIP}
+
+COPY common/ /opt/buildtools/
+COPY requirements-test.txt /opt/buildtools/
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        curl \
+        dos2unix \
+        gcc \
+        g++ \
+        git \
+        make \
+        python3 \
+        python3-dev \
+        python3-pip \
+        python3-venv \
+        tar \
+        tzdata \
+        unzip \
+        vim \
+        wget \
+    && ln -sf /usr/bin/python3 /usr/bin/python \
+    && rm -rf /var/lib/apt/lists/*
+
+# Upgrade pip/setuptools/wheel
+RUN python3 -m pip install --no-cache-dir --upgrade pip setuptools wheel
+
+# Install CANN and OBS
+RUN chmod -R 755 /opt/buildtools/* \
+    && dos2unix /opt/buildtools/* \
+    && /opt/buildtools/install_cann.sh \
+    && /opt/buildtools/install_obs.sh
+
+# Install triton-ascend
+RUN /opt/buildtools/install_triton.sh 3.10
+
+# Install PyTorch from official source, then test requirements from PyPI
+RUN python3 -m pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${PYTORCH_VERSION} \
+    && python3 -m pip install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu -r /opt/buildtools/requirements-test.txt
+
+WORKDIR /home
diff --git a/.github/scripts/detect_changed_patches.sh b/.github/scripts/detect_changed_patches.sh
index 59fb6c8dda..f0738d6a09 100644
--- a/.github/scripts/detect_changed_patches.sh
+++ b/.github/scripts/detect_changed_patches.sh
@@ -73,9 +73,7 @@ while IFS= read -r f; do
             #   test_upstream/test/test_autograd.py.patch → test_autograd.py
             #   test_upstream/test/ao/test_foo.py.patch  → ao/test_foo.py
             #   test_upstream/test/inductor/test_minifer.diff → inductor/test_minifer.py
-            TEST_FILE=$(echo "$f" | sed 's|^test_upstream/test/||; s|\.patch$||; s|\.diff$||')
-            # Ensure .py extension for cases where patch suffix was on bare name (e.g. test_foo.diff)
-            [[ "$TEST_FILE" != *.py ]] && TEST_FILE="${TEST_FILE}.py"
+            TEST_FILE=$(echo "$f" | sed 's|^test_upstream/test/||; s|\.patch$||; s|\.diff$|.py|')
             TEST_PATCHES="${TEST_PATCHES}${f},"
             TEST_FILES="${TEST_FILES}${TEST_FILE},"
             echo "  → test patch: $f → test file: ${TEST_FILE}"
diff --git a/.github/scripts/run_npu_test_shard.py b/.github/scripts/run_npu_test_shard.py
index 15c591cd52..d9abd4f2ed 100644
--- a/.github/scripts/run_npu_test_shard.py
+++ b/.github/scripts/run_npu_test_shard.py
@@ -497,6 +497,7 @@ def run_tests_with_tasks_concurrent(
     max_workers: int,
     result_module,
     quick_test: int = None,
+    max_cases_per_batch: int = 100,
 ) -> Tuple[int, float, List[Dict]]:
     """
     Execute pre-collected test cases with concurrent per-case isolation.
@@ -574,8 +575,8 @@ def run_tests_with_tasks_concurrent(
 
     total_cases = len(tasks)
 
-    # Sort and batch tasks: group same-file cases, max 100 per batch
-    batches = sort_and_batch_tasks(tasks, max_cases_per_batch=100)
+    # Sort and batch tasks: group same-file cases, max max_cases_per_batch per batch
+    batches = sort_and_batch_tasks(tasks, max_cases_per_batch=max_cases_per_batch)
 
     print(f"\n{'=' * 80}", flush=True)
     print(f"Pre-collected cases: {total_cases} cases", flush=True)
@@ -759,6 +760,106 @@ def _build_batch_input_json(
     }
 
 
+# ==============================================================================
+# NPU Task Queue Poisoning & Hardware Error Detection
+# ==============================================================================
+#
+# When an aclnn operator fails fatally (operator bug like aclnnRepeat 0-dim,
+# CANN OOM, or hardware AiCore exception), the NPU device context becomes
+# poisoned. The NPU task queue (NPUQueue.cpp) transitions to CAN_EXIT state:
+# all subsequent operators become silent no-ops that produce garbage data
+# (uninitialized device memory). This poisons every remaining test case in
+# the same worker process.
+#
+# Detection: check each case's output for fatal signatures. The poisoning
+# case itself throws a RuntimeError containing one of these strings. After
+# detection, the worker exits with a special exit code; the parent restarts
+# a fresh worker for remaining cases.
+#
+# Two categories of fatal signatures:
+#   1. "The process exits for this inner error":
+#      NPUQueue ERROR_EXIT throw path (operator bugs, OOM). NOT produced
+#      by hardware errors (UCE/ECC go through deviceErrorMap).
+#   2. Hardware errors (EZ9999, EE9999, vector core exception):
+#      AiCore faults that corrupt device state beyond recovery.
+
+NPU_QUEUE_FATAL_EXIT_CODE = 70
+
+# Layer 1: Fatal error signatures (string matching)
+#
+# Two categories:
+#   A. NPUQueue ERROR_EXIT throw path — produces "The process exits for
+#      this inner error" (covers operator bugs like aclnnRepeat, CANN OOM).
+#   B. deviceErrorMap throw path (NPUQueue.cpp:175-183 ThrowDeviceError) —
+#      produces device error labels for hardware faults. These bypass
+#      ERROR_EXIT and go directly to UCE_EXIT / HBM_ECC_EXIT / etc.,
+#      so they do NOT contain the "process exits" signature.
+#
+# Note: EZ9999 / EE9999 / EZ1009 are CANN runtime error codes that appear
+# in the exception message alongside the deviceErrorMap labels. They are
+# kept for additional robustness but are partially redundant with category B.
+NPU_QUEUE_FATAL_SIGNATURES = [
+    # A. NPUQueue ERROR_EXIT (operator bugs, OOM)
+    "The process exits for this inner error",
+    # B. deviceErrorMap labels (hardware faults, NPUQueue.cpp:175-183)
+    "UCE ERROR",
+    "HBM MULTI BIT ECC ERROR",
+    "SUSPECT MEM ERROR",
+    "HCCS LINK ERROR",
+    "HCCL OP RETRY FAILED",
+    "SUSPECT REMOTE ERROR",
+    # C. CANN runtime error codes (redundant with B, kept for robustness)
+    "EZ9999",
+    "EE9999",
+    "EZ1009",
+]
+
+
+def _check_fatal_npu_error(
+    status: str,
+    message: str,
+    stdout: str,
+    stderr: str,
+) -> bool:
+    """Layer 1: Check if a case result contains a known fatal NPU error signature."""
+    if status not in ("failed", "error"):
+        return False
+    combined = (message or "") + "\n" + (stdout or "") + "\n" + (stderr or "")
+    return any(sig in combined for sig in NPU_QUEUE_FATAL_SIGNATURES)
+
+
+def _check_npu_poisoned() -> bool:
+    """
+    Layer 2: Probe NPU device health by running a trivial computation.
+
+    When the NPU task queue is in CAN_EXIT state (poisoned by a prior fatal
+    error), all operators become silent no-ops — NPUQueue.cpp:573-596 Enqueue()
+    returns without executing, and output tensors contain uninitialized device
+    memory. This probe creates a tensor with a known value and verifies the
+    result. If the queue is poisoned, the result will be garbage. If the
+    device is in error state, the sync will throw.
+
+    This catches poisoning paths that signature matching misses (e.g. new
+    CANN error formats, SUSPECT REMOTE ERROR without EZ9999 prefix, etc.).
+
+    Cost: ~1ms per call. Only invoked on failed/error cases where Layer 1
+    did not match, so zero overhead for passing tests.
+    """
+    try:
+        import torch
+        # torch.ones on NPU calls fill_ -> EXEC_NPU_CMD(aclnnInplaceFillScalar)
+        # -> Enqueue -> no-op if CAN_EXIT. sum() similarly no-op.
+        # .item() triggers MakeSureQueueEmpty (sync), which does NOT throw
+        # in CAN_EXIT state (runtime_error left empty, NPUQueue.cpp:313-315),
+        # so it returns garbage from uninitialized device memory.
+        probe = torch.ones(4, device="npu")
+        return probe.sum().item() != 4.0
+    except Exception:
+        # Any exception (sync timeout, device error, OOM) means the device
+        # context is unhealthy and the worker should be restarted.
+        return True
+
+
 def _execute_worker_batch(
     batch: List[CaseExecutionTask],
     batch_id: int,
@@ -795,6 +896,9 @@ def _execute_worker_batch(
         timeout, verbose, shard, shard_type, npu_device_id,
     )
 
+    unexpected_stdout_dir = report_dir / "unexpected_stdout"
+    unexpected_stdout_dir.mkdir(parents=True, exist_ok=True)
+
     while remaining_cases:
         batch_input["cases"] = [
             {
@@ -828,29 +932,60 @@ def _execute_worker_batch(
 
             last_output_time = monotonic()
 
+            unexpected_log_path = unexpected_stdout_dir / f"batch_{batch_id}.log"
+            unexpected_count = 0
+            unexpected_lock = threading.Lock()
+
             def _read_stdout():
-                nonlocal last_output_time
+                nonlocal last_output_time, unexpected_count
                 if proc.stdout:
                     for line in proc.stdout:
                         last_output_time = monotonic()
-                        line = line.strip()
-                        if not line:
+                        raw_line = line.strip()
+                        if not raw_line:
                             continue
                         try:
-                            case_result = json.loads(line)
+                            case_result = json.loads(raw_line)
                         except json.JSONDecodeError:
                             continue
 
+                        if not isinstance(case_result, dict):
+                            with unexpected_lock:
+                                unexpected_count += 1
+                                count = unexpected_count
+                            ts = datetime.now().isoformat()
+                            json_type = type(case_result).__name__
+                            line_preview = raw_line[:10000]
+                            try:
+                                with open(unexpected_log_path, "a", encoding="utf-8") as uf:
+                                    uf.write(
+                                        f"[{ts}] #{count} type={json_type}"
+                                        f" len={len(raw_line)}\n"
+                                        f"{line_preview}\n"
+                                        f"{'-' * 80}\n"
+                                    )
+                            except OSError:
+                                pass
+                            if count == 1:
+                                print(
+                                    f"  [Batch {batch_id}] Unexpected non-dict JSON line"
+                                    f" (type={json_type}, len={len(raw_line)}),"
+                                    f" full details saved to {unexpected_log_path}",
+                                    flush=True,
+                                )
+                            continue
+
                         nodeid = case_result.get("nodeid", "")
                         status = case_result.get("status", "error")
                         duration = case_result.get("duration", 0.0)
+                        message = case_result.get("message", "")
 
                         full_result = {
                             "nodeid": nodeid,
                             "status": status,
                             "duration": duration,
                             "returncode": int(case_result.get("returncode", 1)),
-                            "message": case_result.get("message", ""),
+                            "message": message,
                             "command": case_result.get("command", ""),
                             "file": case_result.get("file", ""),
                             "case_idx": int(case_result.get("case_idx", 0)),
@@ -864,7 +999,7 @@ def _read_stdout():
                             "nodeid": nodeid,
                             "status": status,
                             "duration": duration,
-                            "message": case_result.get("message", "")[:200],
+                            "message": message[:200],
                         })
                         attempt_completed.add(nodeid)
 
@@ -934,6 +1069,26 @@ def _read_stdout():
                     )
                 continue
 
+            # NPU task queue poisoning or hardware device error: worker
+            # detected the fatal condition and exited cleanly between
+            # cases (os._exit(NPU_QUEUE_FATAL_EXIT_CODE)). Unlike coredump,
+            # no case was in progress — restart for ALL unreported cases
+            # without sacrificing any.
+            if returncode == NPU_QUEUE_FATAL_EXIT_CODE:
+                print(
+                    f"  [Batch {batch_id}] NPU fatal error detected by worker,"
+                    f" restarting for {len(not_reported)} remaining cases...",
+                    flush=True,
+                )
+                remaining_cases = not_reported
+                if remaining_cases:
+                    print(
+                        f"  [Batch {batch_id}] Continuing with "
+                        f"{len(remaining_cases)} remaining cases...",
+                        flush=True,
+                    )
+                continue
+
             if returncode < 0:
                 signal_num = -returncode
                 try:
@@ -1208,6 +1363,34 @@ def _worker_main(worker_input_file: str) -> None:
         # Print JSON line to stdout (parent reads in real-time)
         print(json.dumps(case_result, ensure_ascii=False), flush=True)
 
+        # Detect NPU fatal errors (task queue poisoning or hardware fault).
+        # After a fatal error, the NPU device context is poisoned — all
+        # subsequent ops become silent no-ops producing garbage data. Exit
+        # the worker cleanly between cases (no case in progress) so the
+        # parent can restart a fresh worker for remaining cases.
+        #
+        # Two-layer detection:
+        #   Layer 1: signature matching (fast, zero overhead, known patterns)
+        #   Layer 2: probe computation (~1ms, catches unknown patterns)
+        # Layer 2 only runs if Layer 1 didn't match, and only for
+        # failed/error cases. Zero overhead for passing tests.
+        if status in ("failed", "error"):
+            poisoned = _check_fatal_npu_error(
+                status, message, captured_stdout, captured_stderr
+            )
+            if not poisoned:
+                poisoned = _check_npu_poisoned()
+            if poisoned:
+                print(
+                    f"[{case['case_idx']}] NPU fatal error detected, "
+                    f"exiting worker (code {NPU_QUEUE_FATAL_EXIT_CODE}) "
+                    f"to trigger restart for remaining cases",
+                    flush=True,
+                )
+                sys.stdout.flush()
+                sys.stderr.flush()
+                os._exit(NPU_QUEUE_FATAL_EXIT_CODE)
+
     # Write batch results file as fallback
     results_file = report_dir / f"batch_results_{batch_id}.json"
     try:
@@ -1400,6 +1583,13 @@ def parse_args():
         default=4,
         help="Maximum concurrent workers for regular tests (default: 4). Each worker handles one batch of cases.",
     )
+    parser.add_argument(
+        "--max-cases-per-batch",
+        type=int,
+        default=100,
+        help="Maximum cases per batch (default: 100). Set to 1 for true per-case process isolation "
+             "(each test case runs in its own independent subprocess).",
+    )
     parser.add_argument("--verbose", "-v", action="store_true", help="Verbose output")
     parser.add_argument("--quick-test", type=int, default=None, help="Quick test mode: execute only N cases for fast verification (default: None, run all cases)")
     parser.add_argument("--worker", type=str, default=None, help=argparse.SUPPRESS)
@@ -1560,6 +1750,7 @@ def main():
                 effective_workers,
                 result_module,
                 None,  # quick_test already applied above
+                args.max_cases_per_batch,
             )
             info["per_case_isolation"] = True
             info["concurrent_workers"] = effective_workers
@@ -1680,6 +1871,7 @@ def main():
                 effective_workers,
                 result_module,
                 args.quick_test,
+                args.max_cases_per_batch,
             )
             info["execution_mode"] = "serial" if effective_workers == 1 else "concurrent"
             info["concurrent_workers"] = effective_workers
diff --git a/.github/workflows/_torch-npu-upstream-collect.yml b/.github/workflows/_torch-npu-upstream-collect.yml
index 8efe91a424..8225a4f7e0 100644
--- a/.github/workflows/_torch-npu-upstream-collect.yml
+++ b/.github/workflows/_torch-npu-upstream-collect.yml
@@ -67,7 +67,7 @@ jobs:
     steps:
 
       - name: Setup NPU test environment
-        uses: Ascend/pytorch/.github/actions/setup-npu-test-env@v2.7.1
+        uses: kerer-ai/pytorch/.github/actions/setup-npu-test-env@v2.7.1_image
         with:
           python_version: ${{ inputs.python_version }}
           torch_npu_wheel_artifact: ${{ inputs.torch_npu_wheel_artifact }}
@@ -122,8 +122,8 @@ jobs:
           echo "total_cases=${TOTAL_CASES}" >> $GITHUB_OUTPUT
 
           echo "=== Shard configuration ==="
-          echo "Distributed tests: ${DISTRIBUTED_SHARDS} shards (case-level, serial execution, linux-aarch64-a3-16)"
-          echo "Regular tests: ${REGULAR_SHARDS} shards (case-level, 64 workers, linux-aarch64-a3-16)"
+          echo "Distributed tests: ${DISTRIBUTED_SHARDS} shards (case-level, serial execution, linux-aarch64-a3-8)"
+          echo "Regular tests: ${REGULAR_SHARDS} shards (case-level, 16 workers, linux-aarch64-a3-16)"
           echo "Total cases: ${TOTAL_CASES}"
 
           # Package error logs if any (place at workspace root for flat artifact layout)
diff --git a/.github/workflows/_torch-npu-upstream-report.yml b/.github/workflows/_torch-npu-upstream-report.yml
index 23bf4f62b8..ffa95209af 100644
--- a/.github/workflows/_torch-npu-upstream-report.yml
+++ b/.github/workflows/_torch-npu-upstream-report.yml
@@ -112,7 +112,7 @@ jobs:
             --patch-count "${{ inputs.patch_count }}" \
             --shard-matrix-json "${COMBINED_MATRIX}" \
             --docker-image "${{ inputs.docker_image }}" \
-            --runner "linux-aarch64-a3-16 (distributed, serial), linux-aarch64-a3-16 (regular, 64 workers), linux-aarch64-a3-8 (custom)" \
+            --runner "linux-aarch64-a3-8 (distributed, serial), linux-aarch64-a3-16 (regular, 16 workers), linux-aarch64-a3-16 (custom)" \
             --cases-summary cases-shards/cases_collection_summary.json \
             --cases-by-file-dir cases-shards
 
diff --git a/.github/workflows/_torch-npu-upstream-test-custom.yml b/.github/workflows/_torch-npu-upstream-test-custom.yml
index 3aa1ba2294..f46f7770da 100644
--- a/.github/workflows/_torch-npu-upstream-test-custom.yml
+++ b/.github/workflows/_torch-npu-upstream-test-custom.yml
@@ -40,7 +40,7 @@ jobs:
     steps:
 
       - name: Setup NPU test environment
-        uses: Ascend/pytorch/.github/actions/setup-npu-test-env@v2.7.1
+        uses: kerer-ai/pytorch/.github/actions/setup-npu-test-env@v2.7.1_image
         with:
           python_version: ${{ inputs.python_version }}
           torch_npu_wheel_artifact: ${{ inputs.torch_npu_wheel_artifact }}
@@ -49,25 +49,216 @@ jobs:
 
       - name: Run custom test files
         id: run_tests
+        env:
+          CI: ''
         run: |
           source /usr/local/Ascend/cann/set_env.sh 2>/dev/null || true
           source /usr/local/Ascend/nnal/atb/set_env.sh 2>/dev/null || true
 
           REPORT_DIR=test-reports
-          mkdir -p ${REPORT_DIR}
+          JUNIT_DIR="${REPORT_DIR}/junit_xmls"
+          CASES_LOG_DIR="${REPORT_DIR}/cases_logs"
+          mkdir -p ${REPORT_DIR} ${JUNIT_DIR} ${CASES_LOG_DIR}
+
+          TEST_FILES="${{ inputs.test_files }}"
           set +e
-          # Custom test files: per-case isolation execution
-          python${{ inputs.python_version }} ascend_pytorch/.github/scripts/run_npu_test_shard.py \
-            --test-files "${{ inputs.test_files }}" \
-            --test-dir pytorch-test-src/test \
-            --disabled-testcases pytorch-test-src/test_upstream/disabled_testcases.json \
-            --report-dir ${REPORT_DIR} \
-            --timeout 1200 \
-            --max-workers 16 \
-            --verbose \
-            2>&1 | tee /tmp/test_custom.log
-
-          TEST_STATUS=${PIPESTATUS[0]}
+
+          # ==================================================================
+          # Execution mode detection (checked in priority order):
+          #   1. SAME_PROCESS: prefix → all nodeids in one pytest (control experiment)
+          #   2. Contains "::" → per-case independent pytest processes (serial)
+          #   3. Otherwise → file path mode (run_npu_test_shard.py)
+          # ==================================================================
+
+          if echo "${TEST_FILES}" | grep -q "^SAME_PROCESS:"; then
+            # ================================================================
+            # SAME_PROCESS mode: all cases in a SINGLE pytest invocation.
+            # Control experiment — eliminates BOTH cross-process HBM
+            # contention AND per-case subprocess isolation.
+            # If this PASSES → proves HBM concurrency is the root cause.
+            # If this FAILS → proves pytest.main() batch is the root cause.
+            # ================================================================
+            echo "=== SAME_PROCESS mode: all cases in one pytest process ==="
+
+            # Strip prefix and get nodeids
+            RAW_NODEIDS="${TEST_FILES#SAME_PROCESS:}"
+            ORIG_CWD=$(pwd)
+            REPORT_DIR=$(realpath "${REPORT_DIR}")
+            JUNIT_DIR="${REPORT_DIR}/junit_xmls"
+            CASES_LOG_DIR="${REPORT_DIR}/cases_logs"
+            mkdir -p ${JUNIT_DIR} ${CASES_LOG_DIR}
+
+            # Switch to test directory
+            TEST_SRC_DIR="pytorch-test-src/test"
+            cd "${TEST_SRC_DIR}"
+            echo "Working directory: $(pwd)"
+
+            # Bind to single NPU device (matching original run behavior
+            # where each worker subprocess was pinned to one device).
+            # This prevents the process from spanning multiple devices
+            # and isolates the per-process memory space.
+            export ASCEND_RT_VISIBLE_DEVICES=0
+            echo "NPU device binding: ASCEND_RT_VISIBLE_DEVICES=0"
+
+            # Build pytest nodeid args (strip test/ prefix from each)
+            IFS=',' read -ra NODEIDS <<< "${RAW_NODEIDS}"
+            PARSED_NODEIDS=()
+            for nid in "${NODEIDS[@]}"; do
+              nid=$(echo "${nid}" | xargs)
+              PARSED_NODEIDS+=("${nid#test/}")
+            done
+
+            TOTAL_CASES=${#PARSED_NODEIDS[@]}
+            echo "Running ${TOTAL_CASES} cases in a single pytest process:"
+            for nid in "${PARSED_NODEIDS[@]}"; do
+              echo "  - ${nid}"
+            done
+
+            # Generate combined log filename
+            xml_file="${JUNIT_DIR}/same_process_combined.xml"
+            log_file="${CASES_LOG_DIR}/same_process_combined.log"
+
+            start_time=$(date +%s)
+            python${{ inputs.python_version }} -m pytest \
+              --color=no -ra --tb=short \
+              "${PARSED_NODEIDS[@]}" \
+              --junitxml="${xml_file}" \
+              --timeout=1200 -v \
+              > "${log_file}" 2>&1
+            rc=$?
+            end_time=$(date +%s)
+            duration=$((end_time - start_time))
+
+            if [ "${rc}" -eq 0 ]; then
+              echo "[PASS] All ${TOTAL_CASES} cases passed in one process (${duration}s)"
+              OVERALL_STATUS=0
+            else
+              echo "[FAIL] Some cases failed (rc=${rc}, ${duration}s)"
+              echo "See full log: ${log_file}"
+              tail -80 "${log_file}"
+              OVERALL_STATUS=1
+            fi
+
+            echo ""
+            echo "=============================================="
+            echo "SAME_PROCESS mode complete"
+            echo "  Total cases: ${TOTAL_CASES}"
+            echo "  Duration:    ${duration}s"
+            echo "  Exit code:   ${rc}"
+            echo "  If PASSED → HBM concurrency IS the root cause"
+            echo "  If FAILED → pytest.main() batch IS the root cause"
+            echo "=============================================="
+            TEST_STATUS=${OVERALL_STATUS}
+
+          elif echo "${TEST_FILES}" | grep -q "::"; then
+            # ================================================================
+            # Nodeid mode: direct pytest per-case, serial, independent process
+            # Each nodeid runs in its own pytest invocation (true per-case
+            # process isolation). Ensures no concurrent HBM access across
+            # Dies on the same A3 Chip.
+            # ================================================================
+            echo "=== Nodeid serial mode: each case in its own pytest process ==="
+
+            # Resolve absolute paths before changing directory
+            ORIG_CWD=$(pwd)
+            REPORT_DIR=$(realpath "${REPORT_DIR}")
+            JUNIT_DIR="${REPORT_DIR}/junit_xmls"
+            CASES_LOG_DIR="${REPORT_DIR}/cases_logs"
+            mkdir -p ${JUNIT_DIR} ${CASES_LOG_DIR}
+
+            # Switch to test directory (matching run_npu_test_shard.py behavior)
+            TEST_SRC_DIR="pytorch-test-src/test"
+            cd "${TEST_SRC_DIR}"
+            echo "Working directory: $(pwd)"
+
+            # Bind to single NPU device (matching original run_npu_test_shard.py
+            # behavior where each worker round-robins to one device).
+            export ASCEND_RT_VISIBLE_DEVICES=0
+            echo "NPU device binding: ASCEND_RT_VISIBLE_DEVICES=0"
+
+            IFS=',' read -ra NODEIDS <<< "${TEST_FILES}"
+            TOTAL_CASES=${#NODEIDS[@]}
+            PASSED=0
+            FAILED=0
+            ERRORS=0
+            OVERALL_STATUS=0
+
+            for i in "${!NODEIDS[@]}"; do
+              nodeid=$(echo "${NODEIDS[$i]}" | xargs)
+
+              # Strip "test/" prefix from nodeid to match cwd (pytorch-test-src/test/)
+              # e.g. "test/functorch/test_ops.py::..." -> "functorch/test_ops.py::..."
+              pytest_nodeid="${nodeid#test/}"
+
+              case_idx=$((i + 1))
+
+              # Generate safe filename from nodeid
+              safe_name=$(echo "${nodeid}" | sed 's/[\/:]/_/g' | sed 's/__*/_/g' | sed 's/^_//')
+              xml_file="${JUNIT_DIR}/custom_${case_idx}_${safe_name}.xml"
+              log_file="${CASES_LOG_DIR}/custom_${case_idx}_${safe_name}.log"
+
+              echo ""
+              echo "=== [${case_idx}/${TOTAL_CASES}] ${nodeid} ==="
+
+              start_time=$(date +%s)
+              python${{ inputs.python_version }} -m pytest \
+                --color=no -ra --tb=short \
+                "${pytest_nodeid}" \
+                --junitxml="${xml_file}" \
+                --timeout=1200 -v \
+                > "${log_file}" 2>&1
+              rc=$?
+              end_time=$(date +%s)
+              duration=$((end_time - start_time))
+
+              case "${rc}" in
+                0) status="passed";  PASSED=$((PASSED + 1));   echo "  [PASS] (${duration}s)" ;;
+                1) status="failed";  FAILED=$((FAILED + 1));   echo "  [FAIL] (${duration}s)"; OVERALL_STATUS=1 ;;
+                *) status="error";   ERRORS=$((ERRORS + 1));   echo "  [ERR] rc=${rc} (${duration}s)"; OVERALL_STATUS=1 ;;
+              esac
+
+              # Append to combined log
+              echo "[${case_idx}/${TOTAL_CASES}] ${status^^} ${nodeid} (${duration}s)" >> /tmp/test_custom.log
+
+              # Also dump failure output to combined log for easy viewing
+              if [ "${rc}" -ne 0 ]; then
+                echo "--- FAILURE LOG: ${nodeid} ---" >> /tmp/test_custom.log
+                tail -80 "${log_file}" >> /tmp/test_custom.log
+                echo "--- END FAILURE: ${nodeid} ---" >> /tmp/test_custom.log
+              fi
+            done
+
+            echo ""
+            echo "=============================================="
+            echo "Serial nodeid execution complete"
+            echo "  Total:  ${TOTAL_CASES}"
+            echo "  Passed: ${PASSED}"
+            echo "  Failed: ${FAILED}"
+            echo "  Errors: ${ERRORS}"
+            echo "  Exit code: ${OVERALL_STATUS}"
+            echo "=============================================="
+            TEST_STATUS=${OVERALL_STATUS}
+
+          else
+            # ================================================================
+            # File path mode: use run_npu_test_shard.py for auto case discovery
+            # Parallel execution matching regular workflow: 16 concurrent
+            # workers (one per NPU device), 100 same-file cases per batch.
+            # ================================================================
+            echo "=== File path mode: using run_npu_test_shard.py ==="
+            python${{ inputs.python_version }} ascend_pytorch/.github/scripts/run_npu_test_shard.py \
+              --test-files "${TEST_FILES}" \
+              --test-dir pytorch-test-src/test \
+              --disabled-testcases pytorch-test-src/test_upstream/disabled_testcases.json \
+              --report-dir ${REPORT_DIR} \
+              --timeout 1200 \
+              --max-workers 16 \
+              --verbose \
+              2>&1 | tee /tmp/test_custom.log
+
+            TEST_STATUS=${PIPESTATUS[0]}
+          fi
+
           echo "status=${TEST_STATUS}" >> $GITHUB_OUTPUT
           # Don't exit with test status - let step succeed to allow report generation
 
@@ -81,15 +272,22 @@ jobs:
             echo "Found ${XML_COUNT} XML files"
             tar -czf test-reports/junit_xmls.tar.gz -C test-reports junit_xmls
             rm -rf test-reports/junit_xmls
-            echo "JUnit XMLs compressed"
+            echo "JUnit XMLs compressed: $(ls -lh test-reports/junit_xmls.tar.gz)"
           fi
 
-          # Package failed cases logs into compressed archive
-          if [ -d "test-reports/failed_cases_logs" ]; then
-            echo "=== Compressing failed cases logs ==="
-            tar -czf test-reports/failed_cases_logs.tar.gz -C test-reports failed_cases_logs
-            rm -rf test-reports/failed_cases_logs
-            echo "Failed cases logs compressed"
+          # Package cases logs into compressed archive
+          if [ -d "test-reports/cases_logs" ]; then
+            echo "=== Compressing cases logs ==="
+            LOGS_COUNT=$(find test-reports/cases_logs -type f | wc -l)
+            echo "Found ${LOGS_COUNT} case log files"
+            tar -czf test-reports/cases_logs.tar.gz -C test-reports cases_logs
+            rm -rf test-reports/cases_logs
+            echo "Cases logs compressed: $(ls -lh test-reports/cases_logs.tar.gz)"
+          fi
+
+          # Package shard_cases.json
+          if [ -f "test-reports/shard_custom-1_cases.json" ]; then
+            echo "Cases JSON exists: $(ls -lh test-reports/shard_custom-1_cases.json)"
           fi
 
       - name: Upload test reports
@@ -115,4 +313,25 @@ jobs:
         with:
           name: error-logs-custom
           path: error-logs-custom.tar.gz
+          retention-days: 60
+
+      - name: Compress /root/ascend CANN logs
+        if: always()
+        run: |
+          if [ -d "/root/ascend" ]; then
+            echo "=== Compressing /root/ascend CANN logs ==="
+            FILE_COUNT=$(find /root/ascend -type f | wc -l)
+            echo "Found ${FILE_COUNT} files under /root/ascend"
+            tar -czf ascend-cann-logs-custom.tar.gz -C /root ascend
+            echo "CANN logs compressed: $(ls -lh ascend-cann-logs-custom.tar.gz)"
+          else
+            echo "/root/ascend directory does not exist, skipping"
+          fi
+
+      - name: Upload CANN logs
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: ascend-cann-logs-custom
+          path: ascend-cann-logs-custom.tar.gz
           retention-days: 60
\ No newline at end of file
diff --git a/.github/workflows/_torch-npu-upstream-test-dist.yml b/.github/workflows/_torch-npu-upstream-test-dist.yml
index 0c47d1095b..99518ea80c 100644
--- a/.github/workflows/_torch-npu-upstream-test-dist.yml
+++ b/.github/workflows/_torch-npu-upstream-test-dist.yml
@@ -35,7 +35,7 @@ defaults:
 jobs:
   run_tests:
     name: test_distributed (${{ matrix.shard }}/${{ inputs.distributed_shards }})
-    runs-on: linux-aarch64-a3-16
+    runs-on: linux-aarch64-a3-8
     timeout-minutes: 1800
     container:
       image: ${{ inputs.docker_image }}
@@ -48,7 +48,7 @@ jobs:
     steps:
 
       - name: Setup NPU test environment
-        uses: Ascend/pytorch/.github/actions/setup-npu-test-env@v2.7.1
+        uses: kerer-ai/pytorch/.github/actions/setup-npu-test-env@v2.7.1_image
         with:
           python_version: ${{ inputs.python_version }}
           torch_npu_wheel_artifact: ${{ inputs.torch_npu_wheel_artifact }}
@@ -63,6 +63,8 @@ jobs:
 
       - name: Run distributed shard ${{ matrix.shard }}/${{ inputs.distributed_shards }}
         id: run_test
+        env:
+          CI: ''
         run: |
           source /usr/local/Ascend/cann/set_env.sh 2>/dev/null || true
           source /usr/local/Ascend/nnal/atb/set_env.sh 2>/dev/null || true
@@ -78,7 +80,7 @@ jobs:
 
           echo "=== Distributed Shard ${{ matrix.shard }} (Case-level) ==="
           echo "Total cases: ${TOTAL_CASES}"
-          echo "Runner: linux-aarch64-a3-16 (16-card NPU)"
+          echo "Runner: linux-aarch64-a3-8"
           echo "Execution mode: SERIAL"
 
           # Distributed tests: pre-collected cases, serial execution
@@ -147,4 +149,25 @@ jobs:
         with:
           name: error-logs-dist-${{ matrix.shard }}
           path: error-logs-dist-${{ matrix.shard }}.tar.gz
+          retention-days: 60
+
+      - name: Compress /root/ascend CANN logs
+        if: always()
+        run: |
+          if [ -d "/root/ascend" ]; then
+            echo "=== Compressing /root/ascend CANN logs ==="
+            FILE_COUNT=$(find /root/ascend -type f | wc -l)
+            echo "Found ${FILE_COUNT} files under /root/ascend"
+            tar -czf ascend-cann-logs-dist-${{ matrix.shard }}.tar.gz -C /root ascend
+            echo "CANN logs compressed: $(ls -lh ascend-cann-logs-dist-${{ matrix.shard }}.tar.gz)"
+          else
+            echo "/root/ascend directory does not exist, skipping"
+          fi
+
+      - name: Upload CANN logs
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: ascend-cann-logs-dist-${{ matrix.shard }}
+          path: ascend-cann-logs-dist-${{ matrix.shard }}.tar.gz
           retention-days: 60
\ No newline at end of file
diff --git a/.github/workflows/_torch-npu-upstream-test-regular.yml b/.github/workflows/_torch-npu-upstream-test-regular.yml
index a0859bab3a..d988fce1e6 100644
--- a/.github/workflows/_torch-npu-upstream-test-regular.yml
+++ b/.github/workflows/_torch-npu-upstream-test-regular.yml
@@ -48,7 +48,7 @@ jobs:
     steps:
 
       - name: Setup NPU test environment
-        uses: Ascend/pytorch/.github/actions/setup-npu-test-env@v2.7.1
+        uses: kerer-ai/pytorch/.github/actions/setup-npu-test-env@v2.7.1_image
         with:
           python_version: ${{ inputs.python_version }}
           torch_npu_wheel_artifact: ${{ inputs.torch_npu_wheel_artifact }}
@@ -61,8 +61,18 @@ jobs:
           name: cases-shards
           path: cases-shards
 
+      - name: Debug all environment variables
+        run: |
+          echo "=== All Environment Variables (secrets filtered) ==="
+          env | sort | grep -ivE \
+            'PASSWORD|PASSWD|SECRET|TOKEN|KEY|CREDENTIAL|PRIVATE|ACCESS|SIGNING|AUTH|CERT|ENC(ODE|RYPT)|SALT|NONCE|ACCOUNT|IDENTITY|LICENSE' \
+            || true
+          echo "=== End ==="
+
       - name: Run regular shard ${{ matrix.shard }}/${{ inputs.regular_shards }}
         id: run_test
+        env:
+          CI: ''
         run: |
           source /usr/local/Ascend/cann/set_env.sh 2>/dev/null || true
           source /usr/local/Ascend/nnal/atb/set_env.sh 2>/dev/null || true
@@ -89,7 +99,7 @@ jobs:
             --disabled-testcases pytorch-test-src/test_upstream/disabled_testcases.json \
             --report-dir ${REPORT_DIR} \
             --timeout 1200 \
-            --max-workers 64 \
+            --max-workers 16 \
             --verbose \
             2>&1 | tee /tmp/test_shard_reg_${{ matrix.shard }}.log
 
@@ -150,4 +160,25 @@ jobs:
         with:
           name: error-logs-reg-${{ matrix.shard }}
           path: error-logs-reg-${{ matrix.shard }}.tar.gz
+          retention-days: 60
+
+      - name: Compress /root/ascend CANN logs
+        if: always()
+        run: |
+          if [ -d "/root/ascend" ]; then
+            echo "=== Compressing /root/ascend CANN logs ==="
+            FILE_COUNT=$(find /root/ascend -type f | wc -l)
+            echo "Found ${FILE_COUNT} files under /root/ascend"
+            tar -czf ascend-cann-logs-reg-${{ matrix.shard }}.tar.gz -C /root ascend
+            echo "CANN logs compressed: $(ls -lh ascend-cann-logs-reg-${{ matrix.shard }}.tar.gz)"
+          else
+            echo "/root/ascend directory does not exist, skipping"
+          fi
+
+      - name: Upload CANN logs
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: ascend-cann-logs-reg-${{ matrix.shard }}
+          path: ascend-cann-logs-reg-${{ matrix.shard }}.tar.gz
           retention-days: 60
\ No newline at end of file
diff --git a/.github/workflows/build-docker-images.yml b/.github/workflows/build-docker-images.yml
new file mode 100644
index 0000000000..5d4199a2ff
--- /dev/null
+++ b/.github/workflows/build-docker-images.yml
@@ -0,0 +1,120 @@
+name: Build Docker Images
+
+on:
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: 'Single image tag to build (without timestamp). Leave empty to build all.'
+        required: false
+        type: string
+        default: ''
+  push:
+    paths:
+      - .ci/docker/**
+      - .github/workflows/build-docker-images.yml
+
+env:
+  REGISTRY: quay.io
+  QUAY_ORG: kerer
+  IMAGE_NAME: pytorch
+
+jobs:
+  matrix:
+    runs-on: ubuntu-latest
+    outputs:
+      tags: ${{ steps.set.outputs.tags }}
+    steps:
+      - id: set
+        run: |
+          if [ -n "${{ inputs.tag }}" ]; then
+            TAGS='["${{ inputs.tag }}"]'
+          else
+            TAGS='["torch-npu-builder-x86_64-torch2.7.1","torch-npu-builder-aarch64-torch2.7.1","torch-npu-test-x86_64-cann-a1-py3.10-torch2.7.1","torch-npu-test-x86_64-cann-a2-py3.10-torch2.7.1","torch-npu-test-x86_64-cann-a3-py3.10-torch2.7.1","torch-npu-test-aarch64-cann-a1-py3.10-torch2.7.1","torch-npu-test-aarch64-cann-a2-py3.10-torch2.7.1","torch-npu-test-aarch64-cann-a3-py3.10-torch2.7.1"]'
+          fi
+          echo "tags=${TAGS}" >> $GITHUB_OUTPUT
+
+  build:
+    needs: matrix
+    environment: QUAY_USERNAME
+    permissions:
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        tag: ${{ fromJSON(needs.matrix.outputs.tags) }}
+    runs-on: ${{ contains(matrix.tag, 'x86_64') && 'ubuntu-latest' || 'ubuntu-22.04-arm' }}
+    steps:
+      - name: Free up disk space
+        run: |
+          sudo rm -rf /usr/local/lib/android /opt/ghc /usr/local/share/boost
+          sudo rm -rf /usr/share/dotnet /usr/local/share/powershell
+          sudo rm -rf /opt/hostedtoolcache
+          docker system prune -af
+          sudo apt clean && sudo apt autoremove -y
+          df -h
+
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Log in to Quay.io
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ secrets.QUAY_USERNAME }}
+          password: ${{ secrets.QUAY_PASSWORD }}
+
+      - name: Build and push image
+        run: |
+          TIMESTAMP=$(date -u +%Y%m%d%H%M)
+          cd .ci/docker
+          TIMESTAMP=${TIMESTAMP} ./docker_build.sh ${{ matrix.tag }}
+
+          IMAGE_TAG="${{ matrix.tag }}-${TIMESTAMP}"
+          REMOTE_IMAGE="${{ env.REGISTRY }}/${{ env.QUAY_ORG }}/${{ env.IMAGE_NAME }}:${IMAGE_TAG}"
+          docker tag "${IMAGE_TAG}" "${REMOTE_IMAGE}"
+          docker push "${REMOTE_IMAGE}"
+
+          mkdir -p /tmp/result
+          echo "${REMOTE_IMAGE}" > "/tmp/result/${{ matrix.tag }}.txt"
+          echo "Pushed ${REMOTE_IMAGE}"
+
+      - name: Upload result
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: result-${{ matrix.tag }}
+          path: /tmp/result/${{ matrix.tag }}.txt
+          retention-days: 1
+
+  summary:
+    needs: [matrix, build]
+    runs-on: ubuntu-latest
+    if: always()
+    steps:
+      - name: Download results
+        uses: actions/download-artifact@v4
+        with:
+          pattern: result-*
+          path: /tmp/results
+          merge-multiple: true
+
+      - name: Generate summary
+        run: |
+          echo "## Docker Image Build Summary" >> $GITHUB_STEP_SUMMARY
+          echo "" >> $GITHUB_STEP_SUMMARY
+          echo "| # | Image | Pull Command |" >> $GITHUB_STEP_SUMMARY
+          echo "|---|-------|-------------|" >> $GITHUB_STEP_SUMMARY
+
+          if [ -d /tmp/results ] && [ "$(ls -A /tmp/results 2>/dev/null)" ]; then
+            COUNT=1
+            for f in /tmp/results/*.txt; do
+              IMAGE=$(cat "$f")
+              echo "| ${COUNT} | \`${IMAGE##*:}\` | \`docker pull ${IMAGE}\` |" >> $GITHUB_STEP_SUMMARY
+              COUNT=$((COUNT + 1))
+            done
+          else
+            echo "| - | No images built | - |" >> $GITHUB_STEP_SUMMARY
+          fi
+
+          echo "" >> $GITHUB_STEP_SUMMARY
+          echo "**Registry:** \`${{ env.REGISTRY }}/${{ env.QUAY_ORG }}/${{ env.IMAGE_NAME }}\`" >> $GITHUB_STEP_SUMMARY
diff --git a/test_upstream/torch_env_patch.sh b/test_upstream/torch_env_patch.sh
index 84d879c1c1..cba99307a3 100755
--- a/test_upstream/torch_env_patch.sh
+++ b/test_upstream/torch_env_patch.sh
@@ -102,4 +102,4 @@ done
 echo ""
 echo "========================================"
 echo "All $count patches applied successfully"
-echo "========================================"
\ No newline at end of file
+echo "========================================"