diff --git a/.ci/docker/README.md b/.ci/docker/README.md
new file mode 100644
index 0000000000..0f8388e268
--- /dev/null
+++ b/.ci/docker/README.md
@@ -0,0 +1,49 @@
+# torch-npu CI Docker Images
+
+本目录管理 torch-npu 项目的测试镜像 (test)，当前仅支持 aarch64 架构。
+
+## 镜像类型
+
+| 类型 | 基座 | 用途 |
+|------|------|------|
+| **test** | ubuntu:22.04 | CI 单元测试运行环境，包含 PyTorch nightly、CANN runtime、triton-ascend 和测试框架 |
+
+## 目录结构
+
+```
+.ci/docker/
+├── README.md
+├── 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 A2/A3)
+│   ├── install_triton.sh         # 安装 triton-ascend
+│   └── install_obs.sh            # 安装华为 OBS util
+└── test/
+    └── Dockerfile.aarch64
+```
+
+## 快速构建
+
+```bash
+./docker_build.sh torch-npu-test-aarch64-cann-a2-py3.10-torch-nightly
+./docker_build.sh torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly
+```
+
+## Tag 命名规范
+
+```
+torch-npu-test-aarch64-cann<CHIP>-py<PYTHON_VERSION>-torch-nightly
+```
+
+| 字段 | 可选值 |
+|------|--------|
+| CHIP | A2 (Ascend 910b), A3 |
+| PYTHON_VERSION | 3.10 |
+
+## CANN 芯片映射
+
+| CANN_CHIP | 芯片 | CANN 版本 |
+|-----------|------|----------|
+| A2 | Ascend 910b | 9.1.0-beta.1 |
+| A3 | Ascend A3 | 9.0.0-beta.2 |
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 "${TRITON_URL}"
+echo "triton-ascend installed."
diff --git a/.ci/docker/docker_build.sh b/.ci/docker/docker_build.sh
new file mode 100755
index 0000000000..b32b0d40f7
--- /dev/null
+++ b/.ci/docker/docker_build.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/bash
+# Docker image build script for torch-npu CI test images (aarch64 only).
+#
+# Usage:
+#   ./docker_build.sh <IMAGE_TAG>
+#
+# Examples:
+#   ./docker_build.sh torch-npu-test-aarch64-cann-a2-py3.10-torch-nightly
+#   ./docker_build.sh torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly
+
+set -euo pipefail
+
+BASE_TAG="${1:?Usage: $0 <IMAGE_TAG>}"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+BUILD_CONTEXT="${SCRIPT_DIR}"
+
+case "$BASE_TAG" in
+  torch-npu-test-aarch64-cann-a2-py3.10-torch-nightly)
+    CANN_CHIP=A2
+    PYTHON_VERSION=3.10
+    ;;
+  torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly)
+    CANN_CHIP=A3
+    PYTHON_VERSION=3.10
+    ;;
+  *)
+    echo "ERROR: Unknown image tag: $BASE_TAG"
+    echo ""
+    echo "Supported tags:"
+    echo "  torch-npu-test-aarch64-cann-a2-py3.10-torch-nightly"
+    echo "  torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly"
+    exit 1
+    ;;
+esac
+
+TIMESTAMP="$(date +%Y%m%d%H%M)"
+TAG="${BASE_TAG}-${TIMESTAMP}"
+
+DOCKERFILE="${SCRIPT_DIR}/test/Dockerfile.aarch64"
+
+echo "=== Image Configuration ==="
+echo "  Architecture:   aarch64"
+echo "  CANN Chip:      ${CANN_CHIP}"
+echo "  Python:         ${PYTHON_VERSION}"
+echo "  PyTorch:        nightly"
+echo "  Full Tag:       ${TAG}"
+
+echo "=== Building test image: ${TAG} ==="
+docker build \
+  --build-arg CANN_CHIP="${CANN_CHIP}" \
+  --build-arg PYTHON_VERSION="${PYTHON_VERSION}" \
+  --tag "${TAG}" \
+  --file "${DOCKERFILE}" \
+  "${BUILD_CONTEXT}"
+
+echo "=== Image built successfully: ${TAG} ==="
diff --git a/.ci/docker/requirements-post.txt b/.ci/docker/requirements-post.txt
new file mode 100644
index 0000000000..779d6634e0
--- /dev/null
+++ b/.ci/docker/requirements-post.txt
@@ -0,0 +1,11 @@
+# Python dependencies that require PyTorch to be pre-installed.
+# Install AFTER PyTorch is built, before running tests.
+# Uses PyTorch nightly index to resolve torch-dependent packages.
+
+--index-url https://download.pytorch.org/whl/nightly/cpu
+--extra-index-url https://pypi.org/simple
+-f https://data.pyg.org/whl/torch-2.13.0+cpu.html
+
+torchvision
+torch_geometric==2.5.3
+torch-scatter==2.1.2
diff --git a/.ci/docker/requirements-test-annotated.md b/.ci/docker/requirements-test-annotated.md
new file mode 100644
index 0000000000..596d7aa8b1
--- /dev/null
+++ b/.ci/docker/requirements-test-annotated.md
@@ -0,0 +1,181 @@
+# requirements-test.txt 依赖说明文档
+
+> **用途**: 安装在 torch-npu CI Docker 测试镜像中，用于构建和运行上游 PyTorch 测试用例。
+>
+> **对齐基准**: 上游 PyTorch `.ci/docker/requirements-ci.txt` (py3.10 / jammy profile)
+>
+> **安装位置**: `.ci/docker/test/Dockerfile.aarch64:78`
+> ```dockerfile
+> RUN conda run -n py_${PYTHON_VERSION} pip install --no-cache-dir -r /opt/buildtools/requirements-test.txt
+> ```
+
+---
+
+## 目录
+
+- [Core test frameworks](#core-test-frameworks)
+- [Type checking](#type-checking)
+- [Core scientific](#core-scientific)
+- [Data & serialization](#data--serialization)
+- [ONNX](#onnx)
+- [ML / data processing](#ml--data-processing)
+- [Image / audio / ML](#image--audio--ml)
+- [Utilities](#utilities)
+- [Build tools](#build-tools)
+- [Solver / optimization](#solver--optimization)
+- [Test utilities](#test-utilities)
+- [Infrastructure](#infrastructure)
+
+---
+
+## Core test frameworks
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `pytest` | `7.3.2` | ✅ 一致 | 全部 `test/*.py` | 测试框架主体，`test/run_test.py` 核心调度 |
+| `pytest-xdist` | `3.3.1` | ✅ 一致 | `test/conftest.py:142`, `test/run_test.py:1361` | 并行测试执行，`-n` 参数 |
+| `pytest-flakefinder` | `1.1.0` | ✅ 一致 | `test/run_test.py:547,2289` | 重跑测试以发现 flaky tests |
+| `pytest-rerunfailures` | `>=10.3` | ✅ 一致 | `test/conftest.py:42` | 失败用例自动重试 |
+| `pytest-subtests` | `0.13.1` | ✅ 一致 | `test/distributed/_composable/test_replicate_mixed_precision.py:73` | 子测试支持 |
+| `pytest-cpp` | `2.3.0` | ✅ 一致 | `test/run_test.py:548` | pytest 调用 C++ 测试 |
+| `xdoctest` | `1.3.0` | ✅ 一致 | `test/run_test.py:1162-1265` | 运行 docstring 中的 doctest |
+| `hypothesis` | `6.56.4` | ✅ 一致 | `test/quantization/fx/test_quantize_fx.py:157`, `test/nn/test_nn.py` 等 | 基于属性的测试生成 |
+| `expecttest` | `0.3.0` | ✅ 一致 | `test/test_tensorboard.py:11`, `test/inductor/test_cutedsl_template.py:5`, `test/benchmark_utils/test_benchmark_utils.py:12` 等 | 期望值自动填充测试 |
+| `parameterized` | `0.8.1` | ✅ 一致 | `test/onnx/torchlib/test_ops.py:38`, `test/higher_order_ops/test_invoke_subgraph.py:12`, `test/onnx/test_models_onnxruntime.py:9` 等 | 参数化测试 |
+| `unittest-xml-reporting` | `<=3.2.0,>=2.0.0` | ✅ 一致 | CI 测试结果 XML 输出 | `lxml` 是其依赖 |
+
+## Type checking
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `mypy` | `1.16.0` | ✅ 一致 | `test/test_type_hints.py:22` `import mypy.api`, `test/test_typing.py:22` `from mypy import api` | 类型检查测试 |
+
+## Core scientific
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `numpy` | `1.23.2` | ✅ 一致 (py3.10) | `test/` 下 277 个文件 import | 核心数值计算库 |
+| `scipy` | `1.10.1` | ✅ 一致 (py<=3.11) | `test/` 下 41 个文件 import | 科学计算，`test_linalg.py`, `test_binary_ufuncs.py` 等 |
+| `opt-einsum` | `3.3` | ✅ 一致 | `test/test_linalg.py` (einsum 优化) | 张量缩并顺序优化 |
+| `sympy` | `1.13.3` | ✅ 一致 | `test/` 下 69 个文件引用 | 符号计算，形状推断、guard 条件 |
+| `networkx` | `2.8.8` | ✅ 一致 | `test/functorch/test_aotdispatch.py:134,7088` | 复杂网络库，AOT dispatch 测试 |
+
+## Data & serialization
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `pillow` | `12.3.0` | ✅ 一致 | `test/test_tensorboard.py:118` `from PIL import Image`, `test/onnx/test_models_onnxruntime.py:10` | 图像处理 |
+| `protobuf` | `6.33.5` | ✅ 一致 | `test/test_tensorboard.py:117` `from google.protobuf import text_format`, `test/onnx/torchlib/error_reproduction.py:28` | 序列化协议 |
+| `dill` | `0.3.7` | ✅ 一致 | `test/test_datapipe.py:593` `@skipIfNoDill` | 序列化库，DataLoader 测试 |
+| `flatbuffers` | `24.12.23` | ✅ 一致 | `test/onnx/` 目录 (ONNX 序列化) | 跨平台序列化 |
+| `lxml` | `5.3.0` | ✅ 一致 (py<3.14) | `unittest-xml-reporting` 的依赖 | XML 处理，测试报告 |
+
+## ONNX
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `onnx` | `1.21.0` | ✅ 一致 | `test/` 下 57 个文件 import | ONNX 模型格式 |
+| `onnx-ir` | `0.1.16` | ✅ 一致 | `onnxscript` 的依赖 | ONNX 内存 IR |
+| `onnxruntime` | `1.23.1` | 上游注释掉，通过 `install_onnx.sh` 安装 | **torch-npu**: `test/onnx/test_pytorch_onnx_onnxruntime_npu.py:7`, `test/onnx/dynamo/test_dynamo_with_onnxruntime_backend.py:12`, `test/onnx/onnx_test_common.py:16`, `test/onnx/test_pytorch_jit_onnx.py:2`; **上游**: `test/onnx/` 7 个文件 | ONNX 推理引擎，ONNX 测试必需 |
+| `onnxscript` | `0.6.2` | ✅ 一致 | `test/onnx/exporter/test_tensors.py:6`, `test/onnx/exporter/test_api.py:10`, `test/onnx/test_onnxscript_runtime.py:8` | ONNX Script 语言 |
+
+## ML / data processing
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `jinja2` | `3.1.6` | ✅ 一致 | `torch/_inductor/codegen/common.py:2619,2652`, `torch/_inductor/select_algorithm.py:2506`, `torch/distributed/debug/_frontend.py:15` | 模板引擎，Inductor 代码生成 |
+| `filelock` | `3.20.3` | ✅ 一致 | `test/inductor/test_caching.py:18`, `torch/_inductor/debug.py:482`, `torch/_inductor/runtime/caching/locks.py:18` | 文件锁，Inductor 缓存 |
+| `optree` | `0.13.0` | ✅ 一致 (py<3.14) | `test/` 下 4 个文件 import | 树结构操作，pytree |
+| `ml-dtypes` | `0.5.1` | 上游不存在 | **torch-npu**: `test/custom_ops/test_npu_anti_quant.py:3` `from ml_dtypes import int4`; **上游**: `test/onnx/exporter/test_core.py:10` | 机器学习数据类型，NPU 量化算子测试 |
+| `transformers` | `4.36.2` | 上游通过 `install_onnx.sh` 安装 `4.36.2` | **torch-npu**: `test/dynamo/test_model_output.py:13-16`, `test/onnx/dynamo/test_dynamo_with_onnxruntime_backend.py:398-399`; **上游**: `test/dynamo/test_model_output.py`, `test/inductor/test_padding.py`, `test/onnx/exporter/test_small_models_e2e.py` | HuggingFace Transformers，模型导出测试 |
+
+## Image / audio / ML
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `scikit-image` | `0.22.0` | ✅ 一致 | `test/test_nn.py` (图像处理) | 图像处理库 |
+| `librosa` | `>=0.6.2` | ✅ 一致 (py3.10 不 pin) | `test/test_spectral_ops.py` | 音频分析 |
+| `numba` | `0.57.1` | ✅ 一致 (py3.10) | `test/test_numba_integration.py:19` `import numba.cuda` | JIT 编译 |
+| `tensorboard` | `2.13.0` | ✅ 一致 (py<3.13) | `test/test_tensorboard.py:17-18,119-122` | 可视化工具 |
+| `pywavelets` | `1.4.1` | ✅ 一致 (py<3.12) | `scikit-image` 的依赖 | 小波变换 |
+
+## Utilities
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `tabulate` | `0.9.0` | ✅ 一致 | `test/test_compile_benchmark_util.py:16`, `torch/fx/graph.py:2595`, `torch/ao/quantization/fx/_model_report/model_report_visualizer.py:11` | 表格格式化输出 |
+| `psutil` | 无pin | ✅ 一致 | `test/test_cuda.py:26`, `test/distributed/checkpoint/test_state_dict_stager.py:9`, `test/test_dataloader.py:69`, `test/profiler/test_profiler.py:119` | 系统资源监控 |
+| `fsspec` | `>=0.8.5` | 上游在 `requirements.txt` 中 | `test/distributed/checkpoint/test_fsspec.py:10-11`, `torch/distributed/checkpoint/_fsspec_filesystem.py:11,23` | 文件系统抽象 |
+| `packaging` | `24.2` | ✅ 一致 | `test/` 下 7 个文件 | 包版本管理 |
+| `pyyaml` | `6.0.3` | ✅ 一致 | `test/` 下 9 个文件 | YAML 解析 |
+| `typing-extensions` | `4.12.2` | ✅ 一致 (py<3.14) | `test/` + `torch/` 下 307 处引用 | 类型提示扩展 |
+| `requests` | `2.32.0` | 上游注释掉 | `test/distributed/test_debug.py:13-14`, `test/distributed/test_c10d_fr_hook.py:431,700`, `test/distributed/elastic/test_control_plane.py:194` | HTTP 客户端 |
+| `pybind11` | `3.0.1` | 上游通过 `install_triton.sh` 安装 | **torch-npu C++源码**: `torch_npu/csrc/npu/Graph.cpp:716`, `Stream.cpp:1`, `Event.cpp:3`, `Graph.h:4` 等 56 处 `#include <pybind11/...>`; `CMakeLists.txt:304` | C++/Python 绑定，torch_npu 编译必需 |
+| `tqdm` | `>=4.66.0` | ✅ 一致 | `test/profiler/test_profiler.py:112`, `test/scripts/run_cuda_memcheck.py:23`, `benchmarks/dynamo/*` | 进度条 |
+| `click` | 无pin | ✅ 一致 | `test/cpp/aoti_inference/generate_lowered_cpu.py:3`, `torch/csrc/jit/tensorexpr/scripts/bisect.py:68` | CLI 框架 |
+| `pygments` | `2.20.0` | ✅ 一致 | doctest 语法高亮 | 代码高亮 |
+
+## Build tools
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `cmake` | `3.31.6` | ✅ 一致 | `ci/build.sh`, C++ 扩展测试构建 | 构建系统 |
+| `ninja` | `1.13.0` | ✅ 一致 | `test/run_test.py`, `test_cpp_extensions_aot.py`, `test_determination.py` | 构建加速 |
+| `build` | `1.3.0` | ✅ 一致 | `ci/build.sh` `python -m build --wheel --no-isolation` | Python 构建前端 |
+| `pip` | `26.1.2` | ✅ 一致 | 包安装工具 | pip 自身 |
+| `setuptools` | `78.1.1` | ✅ 一致 | `ci/build.sh` `python setup.py build bdist_wheel` | 构建系统 |
+| `pyzstd` | 无pin | ✅ 一致 | Zstandard 压缩（间接依赖） | 压缩库 |
+| `six` | 无pin | ✅ 一致 | `torch/` 间接依赖 (NNPACK -> PeachPy -> six) | Python 2/3 兼容 |
+| `wheel` | 无pin | 上游在 `requirements.txt` 中 | **torch-npu**: `setup.py:26` `from wheel.bdist_wheel import bdist_wheel`, `setup.py:610,626,631,677` | Wheel 打包，torch_npu `bdist_wheel` 构建必需 |
+
+## Solver / optimization
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `z3-solver` | `4.15.1.0` | ✅ 一致 (非 s390x) | `test/dynamo/test_exc.py:397`, `test/fx/test_z3_gradual_types.py:35`, `test/test_sympy_utils.py:1048`, `test/test_fx_experimental.py:2295` | Z3 定理证明器 |
+| `pulp` | `2.9.0` | ✅ 一致 | `test/distributed/_tools/test_sac_ilp.py` | ILP 求解器 |
+
+## Test utilities
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `coverage` | 无pin | 上游不存在 | `test/run_test.py:2418` `from coverage import Coverage`, `test/run_test.py:506` `["coverage", "run", "--parallel-mode", "--source=torch"]` | 代码覆盖率，`--coverage` 选项 |
+| `junitparser` | `2.1.1` | ✅ 一致 | CI 报告解析 JUnit XML | 测试结果解析 |
+| `lark` | `0.12.0` | ✅ 一致 | `test/` (解析器) | 解析器库 |
+| `pytest-timeout` | `2.3.1` | 上游不存在 | **torch-npu CI**: `run_npu_test_shard.py:1256` `pytest_args.append(f"--timeout={timeout}")`, `run_npu_test_file.py:268` `f"--timeout={case_timeout}"`, `test-npu.sh:55` `PYTEST_ADDOPTS="... -p timeout --timeout=600"`, `npu_poisoning_plugin.py:77` | pytest 超时控制，nightly CI 必需 |
+| `tlparse` | `0.4.0` | ✅ 一致 | CI 日志解析 | 日志解析工具 |
+
+## Infrastructure
+
+| 包名 | 版本 | 上游对应 | 引用位置 | 说明 |
+|------|------|---------|---------|------|
+| `boto3` | `1.35.42` | ✅ 一致 | CI 基础设施（OBS/S3 上传） | AWS SDK |
+| `PyGithub` | `2.3.0` | ✅ 一致 | CI 状态上报 | GitHub API |
+| `redis` | `>=4.0.0` | ✅ 一致 | `test/distributed/tensor/` (OSS 缓存测试), `inductor/test_codecache.py`, `inductor/test_max_autotune.py` | Redis 缓存测试 |
+| `aiohttp` | `3.14.3` | ✅ 一致 | `test/distributed/test_debug.py:34` `import aiohttp`, `torch/distributed/debug/_frontend.py:105` | 异步 HTTP |
+| `dataclasses_json` | `0.6.7` | ✅ 一致 | `tools/stats/` (数据管道) | dataclass JSON 序列化 |
+
+---
+
+## 已删除的包
+
+以下包在本次对齐中从 `requirements-test.txt` 中删除，因为在 torch-npu 仓库、上游 PyTorch 仓库、nightly CI 流程中均无引用：
+
+| 包名 | 原版本 | 删除原因 |
+|------|--------|---------|
+| `beartype` | `0.17.0` | 整个仓库（含 torch-npu 和上游 pytorch）无任何 `import beartype` 或 `@beartype` |
+| `attrs` | 无pin | 无 `import attrs` 或 `from attrs`，是其他包的间接依赖，pip 自动处理 |
+| `decorator` | 无pin | 无 `import decorator` 或 `from decorator`，`test_decorators.py` 是测试 Python decorator 语法而非 `decorator` 包 |
+| `importlib_metadata` | 无pin | 测试使用 stdlib `importlib.metadata`（`torch_npu/contrib/transfer_to_npu.py:5`），非 backport 包 |
+| `zstandard` | `0.25.0` | 整个仓库（含 torch-npu 和上游 pytorch）无任何 `import zstandard`；注意 `pyzstd` 是不同的包 |
+
+## 版本对齐修改
+
+| 包名 | 修改前 | 修改后 | 原因 |
+|------|--------|--------|------|
+| `pillow` | `12.2.0` | `12.3.0` | 与上游 `requirements-ci.txt:172` 对齐 |
+| `pip` | `26.0.1` | `26.1.2` | 与上游 `requirements-ci.txt:379` 对齐 |
+| `aiohttp` | `3.13.4` | `3.14.3` | 与上游 `requirements-ci.txt:429` 对齐 |
+| `typing-extensions` | `4.15.0` | `4.12.2` | 与上游 `requirements-ci.txt:273` 对齐 (py<3.14) |
+| `librosa` | `0.10.2` | `>=0.6.2` | 与上游 `requirements-ci.txt:71` 对齐 (py3.10 不 pin) |
+| `transformers` | `4.40.0` | `4.36.2` | 与上游 `install_onnx.sh:17` 对齐 |
diff --git a/.ci/docker/requirements-test.txt b/.ci/docker/requirements-test.txt
new file mode 100644
index 0000000000..1f3e09ef23
--- /dev/null
+++ b/.ci/docker/requirements-test.txt
@@ -0,0 +1,95 @@
+# Python dependencies required for CI unit tests
+# Aligned with PyTorch upstream .ci/docker/requirements-ci.txt (CUDA 13.0 / py3.10 / jammy)
+# Note: torch-dependent packages (torchvision, torch_geometric, torch-scatter)
+#       are in requirements-post.txt, installed after PyTorch is built.
+
+# Core test frameworks
+pytest==7.3.2
+pytest-xdist==3.3.1
+pytest-flakefinder==1.1.0
+pytest-rerunfailures>=10.3
+pytest-subtests==0.13.1
+pytest-cpp==2.3.0
+xdoctest==1.3.0
+hypothesis==6.56.4
+expecttest==0.3.0
+parameterized==0.8.1
+unittest-xml-reporting<=3.2.0,>=2.0.0
+
+# Type checking
+mypy==1.16.0
+
+# Core scientific
+numpy==1.23.2
+scipy==1.10.1
+opt-einsum==3.3
+sympy==1.13.3
+networkx==2.8.8
+
+# Data & serialization
+pillow==12.3.0
+protobuf==6.33.5
+dill==0.3.7
+flatbuffers==24.12.23
+lxml==5.3.0
+
+# ONNX
+onnx==1.21.0
+onnx-ir==0.1.16
+onnxruntime==1.23.1
+onnxscript==0.6.2
+
+# ML / data processing
+jinja2==3.1.6
+filelock==3.20.3
+optree==0.13.0
+ml-dtypes==0.5.1
+transformers==4.36.2
+
+# Image / audio / ML
+scikit-image==0.22.0
+librosa>=0.6.2
+numba==0.57.1
+tensorboard==2.13.0
+pywavelets==1.4.1
+
+# Utilities
+tabulate==0.9.0
+psutil
+fsspec>=0.8.5
+packaging==24.2
+pyyaml==6.0.3
+typing-extensions==4.12.2
+requests==2.32.0
+pybind11==3.0.1
+tqdm>=4.66.0
+click
+pygments==2.20.0
+
+# Build tools (for C++ extension tests)
+cmake==3.31.6
+ninja==1.13.0
+build==1.3.0
+pip==26.1.2
+setuptools==78.1.1
+pyzstd
+six
+wheel
+
+# Solver / optimization
+z3-solver==4.15.1.0
+pulp==2.9.0
+
+# Test utilities
+coverage
+junitparser==2.1.1
+lark==0.12.0
+pytest-timeout==2.3.1
+tlparse==0.4.0
+
+# Infrastructure
+boto3==1.35.42
+PyGithub==2.3.0
+redis>=4.0.0
+aiohttp==3.14.3
+dataclasses_json==0.6.7
diff --git a/.ci/docker/test/Dockerfile.aarch64 b/.ci/docker/test/Dockerfile.aarch64
new file mode 100644
index 0000000000..e99bf34918
--- /dev/null
+++ b/.ci/docker/test/Dockerfile.aarch64
@@ -0,0 +1,81 @@
+FROM ubuntu:22.04
+
+ARG CANN_CHIP=A2
+ARG PYTHON_VERSION=3.10
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV TZ=Asia/Shanghai
+ENV PATH=/usr/local/bin:$PATH
+ENV CANN_CHIP=${CANN_CHIP}
+
+# Conda environment (aligned with upstream install_conda.sh)
+ENV ANACONDA_PYTHON_VERSION=${PYTHON_VERSION}
+ENV PATH=/opt/conda/envs/py_${PYTHON_VERSION}/bin:/opt/conda/bin:$PATH
+
+COPY common/ /opt/buildtools/
+COPY requirements-test.txt /opt/buildtools/
+
+# System dependencies (aligned with upstream install_base.sh for jammy)
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        autoconf \
+        automake \
+        build-essential \
+        ca-certificates \
+        cmake=3.22* \
+        curl \
+        dos2unix \
+        gcc \
+        g++ \
+        gdb \
+        gfortran \
+        git \
+        jq \
+        libasound2-dev \
+        libffi-dev \
+        libbz2-dev \
+        libgl1 \
+        libjemalloc2 \
+        libjpeg-dev \
+        libreadline-dev \
+        libsndfile-dev \
+        libsqlite3-dev \
+        libssl-dev \
+        libtool \
+        libyaml-dev \
+        libz-dev \
+        ccache \
+        make \
+        sudo \
+        tar \
+        tzdata \
+        unzip \
+        valgrind \
+        vim \
+        wget \
+        zip \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Miniforge3 and create conda env (aligned with upstream install_conda.sh)
+RUN curl -fsSL https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-$(uname -m).sh -o /tmp/miniforge.sh \
+    && bash /tmp/miniforge.sh -b -p /opt/conda \
+    && rm /tmp/miniforge.sh \
+    && conda create -n py_${PYTHON_VERSION} -y \
+        python=${PYTHON_VERSION} \
+        pip \
+        cmake=3.31.6 \
+    && conda clean -ya
+
+# 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 ${PYTHON_VERSION}
+
+# Install base test requirements (no PyTorch-dependent packages; torch is built separately)
+RUN conda run -n py_${PYTHON_VERSION} pip install --no-cache-dir -r /opt/buildtools/requirements-test.txt
+
+WORKDIR /home
diff --git a/.ci/pytorch/build.sh b/.ci/pytorch/build.sh
new file mode 100755
index 0000000000..9ce6b2d6a5
--- /dev/null
+++ b/.ci/pytorch/build.sh
@@ -0,0 +1,76 @@
+#!/usr/bin/bash
+# Main CI build entry point for torch-npu.
+# Modeled after PyTorch upstream .ci/pytorch/build.sh
+#
+# Orchestrates:
+#   1. PyTorch build (from source or pre-built wheel)
+#   2. torch_npu build
+#   3. Integration verification
+#
+# Environment variables:
+#   PYTORCH_SRC          - Path to PyTorch source (default: ../pytorch)
+#   TORCH_NPU_SRC        - Path to torch_npu source (default: repo root)
+#   BUILD_MODE           - "develop" (PR) or "wheel" (nightly)
+#   PYTORCH_BUILD_MODE   - "source" (build from source) or "wheel" (install pre-built)
+#   PYTORCH_WHEEL_URL    - URL/path to pre-built PyTorch wheel (when PYTORCH_BUILD_MODE=wheel)
+#   PYTORCH_VERSION      - PyTorch version for pre-built wheel (default: 2.7.1)
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/common.sh"
+
+PYTORCH_SRC="${PYTORCH_SRC:-${REPO_ROOT}/../pytorch}"
+TORCH_NPU_SRC="${TORCH_NPU_SRC:-${REPO_ROOT}}"
+BUILD_MODE="${BUILD_MODE:-develop}"
+PYTORCH_BUILD_MODE="${PYTORCH_BUILD_MODE:-source}"
+PYTORCH_VERSION="${PYTORCH_VERSION:-2.7.1}"
+PYTORCH_INDEX_URL="${PYTORCH_INDEX_URL:-https://download.pytorch.org/whl/cpu}"
+
+echo "============================================"
+echo "  torch-npu CI Build"
+echo "============================================"
+echo "  Build mode:       ${BUILD_MODE}"
+echo "  PyTorch mode:     ${PYTORCH_BUILD_MODE}"
+echo "  PyTorch src:      ${PYTORCH_SRC}"
+echo "  torch_npu src:    ${TORCH_NPU_SRC}"
+echo "============================================"
+
+# Step 1: Install or build PyTorch
+case "${PYTORCH_BUILD_MODE}" in
+  source)
+    echo ">>> Step 1/3: Building PyTorch from source..."
+    PYTORCH_SRC="${PYTORCH_SRC}" BUILD_MODE="${BUILD_MODE}" bash "${SCRIPT_DIR}/build_pytorch.sh"
+    ;;
+  wheel)
+    echo ">>> Step 1/3: Installing pre-built PyTorch ${PYTORCH_VERSION}..."
+    pip_install --index-url "${PYTORCH_INDEX_URL}" "torch==${PYTORCH_VERSION}"
+    conda_run python -c "import torch; print(f'PyTorch {torch.__version__} installed')"
+    ;;
+  *)
+    echo "ERROR: Unknown PYTORCH_BUILD_MODE: ${PYTORCH_BUILD_MODE}"
+    exit 1
+    ;;
+esac
+
+# Step 2: Build torch_npu
+echo ">>> Step 2/4: Building torch_npu..."
+TORCH_NPU_SRC="${TORCH_NPU_SRC}" BUILD_MODE="${BUILD_MODE}" bash "${SCRIPT_DIR}/build_torch_npu.sh"
+
+# Step 3: Install post-build dependencies (packages that require PyTorch)
+echo ">>> Step 3/4: Installing post-build dependencies..."
+REQUIREMENTS_POST="${TORCH_NPU_SRC}/.ci/docker/requirements-post.txt"
+if [ -f "${REQUIREMENTS_POST}" ]; then
+  pip_install -r "${REQUIREMENTS_POST}"
+  echo "Post-build dependencies installed."
+else
+  echo "WARNING: requirements-post.txt not found at ${REQUIREMENTS_POST}, skipping."
+fi
+
+# Step 4: Integration verification
+echo ">>> Step 4/4: Verifying integration..."
+bash "${SCRIPT_DIR}/integration_verify.sh"
+
+echo "============================================"
+echo "  Build Summary: SUCCESS"
+echo "============================================"
diff --git a/.ci/pytorch/build_pytorch.sh b/.ci/pytorch/build_pytorch.sh
new file mode 100755
index 0000000000..f4c054d2c2
--- /dev/null
+++ b/.ci/pytorch/build_pytorch.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/bash
+# Build PyTorch from source for NPU CI.
+# Modeled after PyTorch upstream .ci/pytorch/build.sh
+#
+# This mirrors the upstream CPU-only build path:
+#   USE_CUDA=0 USE_XNNPACK=0 WERROR=1 python -m build --wheel --no-isolation
+#
+# For CI PR verification, we use setup.py develop (faster, no packaging).
+# For nightly/release, we use python -m build --wheel (full verification).
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/common.sh"
+
+PYTORCH_SRC="${PYTORCH_SRC:-${REPO_ROOT}/../pytorch}"
+BUILD_MODE="${BUILD_MODE:-develop}"  # "develop" for PR, "wheel" for nightly
+
+setup_build_env
+
+echo "=== Building PyTorch from source ==="
+echo "  Source dir:  ${PYTORCH_SRC}"
+echo "  Build mode:  ${BUILD_MODE}"
+
+cd "${PYTORCH_SRC}"
+
+# Install PyTorch build requirements
+echo "=== Installing PyTorch build requirements ==="
+pip_install -r requirements.txt
+
+# Install numpy for build compatibility
+pip_install "numpy>=1.23,<3.0"
+
+case "${BUILD_MODE}" in
+  develop)
+    echo "=== Building PyTorch (develop mode) ==="
+    conda_run python setup.py clean
+    conda_run python setup.py develop
+    echo "=== PyTorch develop install complete ==="
+    ;;
+  wheel)
+    echo "=== Building PyTorch (wheel mode) ==="
+    conda_run python setup.py clean
+    conda_run python -m build --wheel --no-isolation
+    pip_install_whl "$(echo dist/*.whl)"
+    echo "=== PyTorch wheel build and install complete ==="
+    ;;
+  *)
+    echo "ERROR: Unknown build mode: ${BUILD_MODE}"
+    exit 1
+    ;;
+esac
+
+# Verify PyTorch installation
+echo "=== Verifying PyTorch installation ==="
+conda_run python -c "
+import torch
+print(f'PyTorch version: {torch.__version__}')
+print(f'CUDA available: {torch.cuda.is_available()}')
+print(f'Build passed: OK')
+"
+
+echo "=== PyTorch build successful ==="
diff --git a/.ci/pytorch/build_torch_npu.sh b/.ci/pytorch/build_torch_npu.sh
new file mode 100755
index 0000000000..e2e8cc6bfb
--- /dev/null
+++ b/.ci/pytorch/build_torch_npu.sh
@@ -0,0 +1,57 @@
+#!/usr/bin/bash
+# Build torch_npu from source for NPU CI.
+# Must be run AFTER PyTorch is installed (via build_pytorch.sh or pre-built wheel).
+#
+# Usage:
+#   ./build_torch_npu.sh [--python=3.10] [--mode=develop|wheel]
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/common.sh"
+
+TORCH_NPU_SRC="${TORCH_NPU_SRC:-${REPO_ROOT}}"
+PYTHON_VERSION="${ANACONDA_PYTHON_VERSION:-3.10}"
+BUILD_MODE="${BUILD_MODE:-wheel}"
+
+setup_build_env
+
+echo "=== Building torch_npu from source ==="
+echo "  Source dir:      ${TORCH_NPU_SRC}"
+echo "  Python version:  ${PYTHON_VERSION}"
+echo "  Build mode:      ${BUILD_MODE}"
+
+cd "${TORCH_NPU_SRC}"
+
+# Install torch_npu build requirements
+if [[ -f requirements.txt ]]; then
+  echo "=== Installing torch_npu build requirements ==="
+  pip_install -r requirements.txt
+fi
+
+# Verify PyTorch is installed before building torch_npu
+echo "=== Checking PyTorch installation ==="
+if ! conda_run python -c "import torch; print(torch.__version__)" 2>/dev/null; then
+  echo "ERROR: PyTorch is not installed. Please run build_pytorch.sh first."
+  exit 1
+fi
+
+case "${BUILD_MODE}" in
+  develop)
+    echo "=== Building torch_npu (develop mode) ==="
+    conda_run python setup.py develop
+    echo "=== torch_npu develop install complete ==="
+    ;;
+  wheel)
+    echo "=== Building torch_npu (wheel mode) ==="
+    bash ci/build.sh --python="${PYTHON_VERSION}"
+    pip_install_whl "$(echo dist/torch_npu*.whl)"
+    echo "=== torch_npu wheel build and install complete ==="
+    ;;
+  *)
+    echo "ERROR: Unknown build mode: ${BUILD_MODE}"
+    exit 1
+    ;;
+esac
+
+echo "=== torch_npu build successful ==="
diff --git a/.ci/pytorch/common.sh b/.ci/pytorch/common.sh
new file mode 100755
index 0000000000..bf82721289
--- /dev/null
+++ b/.ci/pytorch/common.sh
@@ -0,0 +1,67 @@
+#!/usr/bin/bash
+# Shared utilities for torch-npu CI build scripts.
+# Modeled after PyTorch upstream .ci/pytorch/common.sh and common-build.sh.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+
+# --- Conda helpers (aligned with upstream common_utils.sh) ---
+
+conda_run() {
+  conda run -n "py_${ANACONDA_PYTHON_VERSION:-3.10}" --no-capture-output "$@"
+}
+
+pip_install() {
+  conda_run pip install --progress-bar off "$@"
+}
+
+pip_install_whl() {
+  conda_run pip install --progress-bar off "$@"
+}
+
+# --- Build environment setup ---
+
+setup_build_env() {
+  # Default Python version if not set
+  export ANACONDA_PYTHON_VERSION="${ANACONDA_PYTHON_VERSION:-3.10}"
+
+  # NPU build: no CUDA, no XNNPACK (aligned with upstream CPU-only builds)
+  export USE_CUDA=0
+  export USE_XNNPACK=0
+
+  # Use C++11 ABI (required for torch_npu compatibility)
+  export _GLIBCXX_USE_CXX11_ABI="${_GLIBCXX_USE_CXX11_ABI:-1}"
+
+  # Set MAX_JOBS to avoid OOM
+  if [[ -z "${MAX_JOBS:-}" ]]; then
+    export MAX_JOBS=$(($(nproc) - 2))
+    # Use at least 1 job
+    if [[ "$MAX_JOBS" -lt 1 ]]; then
+      export MAX_JOBS=1
+    fi
+  fi
+
+  echo "=== Build Environment ==="
+  echo "  ANACONDA_PYTHON_VERSION: ${ANACONDA_PYTHON_VERSION}"
+  echo "  USE_CUDA:                ${USE_CUDA}"
+  echo "  USE_XNNPACK:             ${USE_XNNPACK}"
+  echo "  _GLIBCXX_USE_CXX11_ABI:  ${_GLIBCXX_USE_CXX11_ABI}"
+  echo "  MAX_JOBS:                ${MAX_JOBS}"
+  echo "  Python:                  $(conda_run python --version 2>&1)"
+  echo "  Python path:             $(conda_run which python)"
+}
+
+# --- Error handling ---
+
+trap_add() {
+  local trap_add_cmd="$1"
+  shift || true
+  local trap_add_signal="$*"
+
+  if [[ -z "$(trap -p "${trap_add_signal}")" ]]; then
+    trap "${trap_add_cmd}" "${trap_add_signal}"
+  else
+    trap "${trap_add_cmd};$(trap -p "${trap_add_signal}" | sed "s/^trap -- '\(.*\)' \([A-Z0-9 ]*\)$/\1/")" "${trap_add_signal}"
+  fi
+}
diff --git a/.ci/pytorch/integration_verify.sh b/.ci/pytorch/integration_verify.sh
new file mode 100755
index 0000000000..823024f85f
--- /dev/null
+++ b/.ci/pytorch/integration_verify.sh
@@ -0,0 +1,88 @@
+#!/usr/bin/bash
+# Integration verification: confirms PyTorch + torch_npu work together.
+# Tests basic import, device detection, and tensor operations.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/common.sh"
+
+echo "=== Integration Verification ==="
+
+# Test 1: Basic imports
+echo "--- Test 1: Basic imports ---"
+conda_run python -c "
+import torch
+import torch_npu
+print(f'PyTorch: {torch.__version__}')
+print(f'torch_npu: {torch_npu.__version__}')
+print('Import OK')
+"
+
+# Test 2: NPU device detection
+echo "--- Test 2: NPU device detection ---"
+conda_run python -c "
+import torch
+import torch_npu
+count = torch.npu.device_count()
+print(f'NPU device count: {count}')
+print(f'NPU available: {torch.npu.is_available()}')
+if count > 0:
+    print(f'Device name: {torch.npu.get_device_name(0)}')
+print('Device detection OK')
+"
+
+# Test 3: Basic tensor operations on NPU
+echo "--- Test 3: Basic tensor operations ---"
+conda_run python -c "
+import torch
+import torch_npu
+
+if torch.npu.is_available():
+    # Create tensor on NPU
+    x = torch.randn(3, 3).npu()
+    y = torch.randn(3, 3).npu()
+    z = x + y
+    print(f'Tensor addition: {z.mean().item():.4f}')
+
+    # Matmul
+    z = torch.mm(x, y.t())
+    print(f'Matmul shape: {z.shape}')
+
+    # Move back to CPU
+    z_cpu = z.cpu()
+    print(f'CPU tensor: {z_cpu}')
+else:
+    # NPU not available, test CPU fallback
+    x = torch.randn(3, 3)
+    y = torch.randn(3, 3)
+    z = x + y
+    print(f'CPU fallback: {z.mean().item():.4f}')
+    print('(NPU not available, running CPU-only verification)')
+
+print('Tensor operations OK')
+"
+
+# Test 4: Check torch_npu compile status
+echo "--- Test 4: Compilation status ---"
+conda_run python -c "
+import torch
+import torch_npu
+
+# Check that key modules are available
+modules_to_check = [
+    'torch_npu.npu',
+    'torch_npu.optim',
+    'torch_npu.profiler',
+]
+for mod_name in modules_to_check:
+    try:
+        __import__(mod_name)
+        print(f'  {mod_name}: OK')
+    except ImportError:
+        print(f'  {mod_name}: NOT FOUND (may be expected)')
+
+print('Module check complete')
+"
+
+echo "=== Integration verification: ALL PASSED ==="
diff --git a/.github/workflows/_build.yml b/.github/workflows/_build.yml
index 55a0b52dcf..a1fec81053 100644
--- a/.github/workflows/_build.yml
+++ b/.github/workflows/_build.yml
@@ -20,23 +20,52 @@ on:
         required: true
         type: string
         description: The downstream repository full name (owner/repo)
+      downstream_ref:
+        required: false
+        type: string
+        default: 'master'
+        description: The downstream repository branch, tag, or SHA to build
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    runs-on: linux-aarch64-a3-16
+    defaults:
+      run:
+        shell: bash
+    container:
+      image: quay.io/kerer/pytorch:torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly
+      options: --network host
+    env:
+      MAX_JOBS: 128
+      USE_CUDA: 0
+      USE_XNNPACK: 0
+      _GLIBCXX_USE_CXX11_ABI: 1
+      BUILD_IGNORE_SVE_UNAVAILABLE: 1
+      BUILD_WITHOUT_SHA: 1
+      PIP_INDEX_URL: https://mirrors.huaweicloud.com/repository/pypi/simple
     steps:
-      # Step 1: Checkout upstream PyTorch PR code
-      # Use fork repo if available (PR from fork), otherwise use upstream repo (push or PR from same repo)
+      # Step 1: Checkout upstream PyTorch PR code (shallow)
+      # fetch-depth: 1 = single commit only, no branches/tags/history.
+      # PyTorch repo has thousands of branches; full clone takes ~3min.
+      # Submodules init deferred to separate step with --depth=1.
       - name: Checkout upstream PyTorch PR
         uses: actions/checkout@v6
         with:
           repository: ${{ inputs.upstream_fork_repo != '' && inputs.upstream_fork_repo || inputs.upstream_repo }}
           ref: ${{ inputs.upstream_sha }}
-          submodules: recursive
+          fetch-depth: 1
+          submodules: false
           path: pytorch
 
-      # Step 2: Build PyTorch
-      - name: Build PyTorch
+      - name: Init submodules (shallow)
+        working-directory: pytorch
+        run: |
+          git submodule update --init --recursive --depth=1 --jobs=$(nproc)
+
+      # Step 2: Build and install PyTorch
+      - name: Build and install PyTorch
+        id: build_pytorch
+        working-directory: pytorch
         run: |
           echo "============================================"
           echo "Building PyTorch from upstream PR..."
@@ -46,61 +75,186 @@ jobs:
             echo "Repository: ${{ inputs.upstream_repo }}"
           fi
           echo "Commit SHA: ${{ inputs.upstream_sha }}"
+          echo "MAX_JOBS: ${MAX_JOBS}"
           echo "============================================"
-          # TODO: Add actual PyTorch build commands
-          # Example:
-          # cd pytorch
-          # pip install -r requirements.txt
-          # python setup.py develop
-          echo "[SIMULATED] PyTorch build completed successfully!"
-
-      # Step 3: Install PyTorch wheel
-      - name: Install PyTorch wheel
-        run: |
-          echo "============================================"
-          echo "Installing PyTorch wheel..."
-          echo "============================================"
-          # TODO: Add actual install commands
-          # Example:
-          # pip install pytorch/dist/torch*.whl
-          echo "[SIMULATED] PyTorch wheel installed successfully!"
 
-      # Step 4: Checkout downstream repo (torch_npu) master
+          pip install -r requirements.txt
+
+          # Build wheel first, then install from it.
+          # This produces an artifact we can upload for downstream use.
+          mkdir -p /tmp/wheels
+          python -m pip wheel --no-build-isolation -w /tmp/wheels/ . 2>&1 | tee /tmp/build-pytorch.log
+          BUILD_STATUS=${PIPESTATUS[0]}
+
+          if [ ${BUILD_STATUS} -ne 0 ]; then
+            echo "PyTorch build failed!"
+            exit ${BUILD_STATUS}
+          fi
+
+          pip install /tmp/wheels/torch*.whl
+
+          # Must NOT run from pytorch/ source dir: Python would find
+          # the local torch/ source tree instead of the installed package.
+          cd /tmp
+          python -c "import torch; print(f'PyTorch version: {torch.__version__}')"
+          cd "${GITHUB_WORKSPACE}"
+
+      # Step 3: Checkout downstream torch_npu
       - name: Checkout downstream repo (torch_npu)
         uses: actions/checkout@v6
         with:
           repository: ${{ inputs.downstream_repo }}
-          ref: master
-          submodules: recursive
+          ref: ${{ inputs.downstream_ref }}
+          fetch-depth: 1
+          submodules: false
           path: torch_npu
 
-      # Step 5: Build torch_npu
-      - name: Build torch_npu
+      - name: Init submodules (shallow)
+        working-directory: torch_npu
+        run: |
+          git submodule update --init --recursive --depth=1 --jobs=$(nproc)
+
+      # Step 3.5: Diagnose installed PyTorch for Dimname references
+      - name: "[Diagnostic] Check installed torchgen for Dimname"
+        working-directory: /tmp
+        run: |
+          echo "=== Checking installed torchgen for Dimname ==="
+          python -c "
+          import torchgen, os, sys
+          d = os.path.dirname(torchgen.__file__)
+          yaml_path = os.path.join(d, 'packaged/ATen/native/native_functions.yaml')
+          print(f'Checking: {yaml_path}')
+          if not os.path.exists(yaml_path):
+              print('ERROR: native_functions.yaml not found!')
+              sys.exit(1)
+          with open(yaml_path) as f:
+              c = f.read()
+          cnt = c.count('Dimname')
+          print(f'native_functions.yaml: {cnt} Dimname reference(s)')
+          if cnt > 0:
+              for i, l in enumerate(c.split('\n')):
+                  if 'Dimname' in l:
+                      print(f'  line {i}: {l.strip()[:150]}')
+              print('FATAL: Stale native_functions.yaml with Dimname detected!')
+              sys.exit(1)
+          else:
+              print('OK: native_functions.yaml is clean')
+          "
+
+          echo ""
+          echo "=== Checking model.py BaseTy ==="
+          python -c "
+          import torchgen.model as m
+          has = 'Dimname' in [e.name for e in m.BaseTy]
+          print(f'model.py BaseTy has Dimname member: {has}')
+          import sys
+          if has:
+              print('FATAL: model.py still has Dimname in BaseTy!')
+              sys.exit(1)
+          else:
+              print('OK: model.py BaseTy is clean')
+          "
+
+          echo ""
+          echo "=== Checking tags.yaml ==="
+          python -c "
+          import torchgen, os
+          d = os.path.dirname(torchgen.__file__)
+          tags_path = os.path.join(d, 'packaged/ATen/native/tags.yaml')
+          print(f'Checking: {tags_path}')
+          with open(tags_path) as f:
+              c = f.read()
+          cnt = c.count('Dimname')
+          print(f'tags.yaml: {cnt} Dimname reference(s)')
+          if cnt > 0:
+              print('WARNING: tags.yaml has Dimname refs!')
+          else:
+              print('OK: tags.yaml is clean')
+          "
+
+      # Step 4: Build and install torch_npu
+      # NOTE: Intentionally do NOT use 'pip install -r requirements.txt' from torch_npu,
+      # because that file pins a specific torch version which would overwrite the
+      # source-built PyTorch from the upstream PR.
+      - name: Build and install torch_npu
+        id: build_torchnpu
+        working-directory: torch_npu
         run: |
           echo "============================================"
-          echo "Building torch_npu from downstream master..."
+          echo "Building torch_npu..."
           echo "Repository: ${{ inputs.downstream_repo }}"
-          echo "Branch: master"
+          echo "Ref: ${{ inputs.downstream_ref }}"
           echo "============================================"
-          # TODO: Add actual torch_npu build commands
-          # Example:
-          # cd torch_npu
-          # pip install -r requirements.txt
-          # bash ci/build.sh
-          echo "[SIMULATED] torch_npu build completed successfully!"
 
+          bash ci/build.sh --python=3.10 --disable_torchair --disable_rpc 2>&1 | tee /tmp/build-torchnpu.log
+          BUILD_STATUS=${PIPESTATUS[0]}
+
+          if [ ${BUILD_STATUS} -ne 0 ]; then
+            echo "torch_npu build failed!"
+            exit ${BUILD_STATUS}
+          fi
+
+          pip install dist/torch_npu*.whl
+          cp dist/torch_npu*.whl /tmp/wheels/
+
+      # Step 5: Upload wheel artifacts
+      - name: Upload wheel artifacts
+        if: success()
+        uses: actions/upload-artifact@v4
+        with:
+          name: wheels-${{ github.run_number }}
+          path: /tmp/wheels/*.whl
+          retention-days: 7
+
+      # Step 6: Verify NPU environment and imports
+      - name: Verify NPU device
+        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
+
+          echo "=== NPU Device Information ==="
+          npu-smi info
+          echo "=== End of NPU Device Information ==="
+
+      - name: Verify NPU availability
+        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
+
+          python -c "
+          import torch
+          print(f'torch: {torch.__version__}')
+          import torch_npu
+          print(f'torch_npu: {torch_npu.__version__}')
+          print(f'NPU available: {torch.npu.is_available()}')
+          print(f'NPU count: {torch.npu.device_count()}')
+          "
+
+      # Step 7: Upload build artifacts (always, for debugging failures)
+      - name: Upload build artifacts
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: build-artifacts-${{ github.run_number }}
+          path: |
+            /tmp/build-pytorch.log
+            /tmp/build-torchnpu.log
+            torch_npu/torch_npu/csrc/inductor/aoti_torch/generated/c_shim_npu.h
+          if-no-files-found: warn
+
+      # Step 8: Summary
       - name: Summary
+        if: always()
         run: |
           echo "============================================"
-          echo "Build Summary"
+          echo "             Build Summary"
           echo "============================================"
           if [ -n "${{ inputs.upstream_fork_repo }}" ]; then
-            echo "1. PyTorch fork PR: ${{ inputs.upstream_fork_repo }}@${{ inputs.upstream_sha }}"
+            echo "PyTorch fork PR : ${{ inputs.upstream_fork_repo }}"
           else
-            echo "1. PyTorch upstream: ${{ inputs.upstream_repo }}@${{ inputs.upstream_sha }}"
+            echo "PyTorch upstream: ${{ inputs.upstream_repo }}"
           fi
-          echo "2. PyTorch build: SUCCESS (simulated)"
-          echo "3. PyTorch install: SUCCESS (simulated)"
-          echo "4. torch_npu master: ${{ inputs.downstream_repo }}"
-          echo "5. torch_npu build: SUCCESS (simulated)"
-          echo "============================================"
\ No newline at end of file
+          echo "PyTorch commit  : ${{ inputs.upstream_sha }}"
+          echo "torch_npu repo  : ${{ inputs.downstream_repo }}"
+          echo "Max jobs        : ${MAX_JOBS}"
+          echo "============================================"
diff --git a/.github/workflows/build-docker-images.yml b/.github/workflows/build-docker-images.yml
new file mode 100644
index 0000000000..e92a7bdc5b
--- /dev/null
+++ b/.github/workflows/build-docker-images.yml
@@ -0,0 +1,160 @@
+name: Build Docker Images
+
+on:
+  push:
+    paths:
+      - '.ci/docker/**'
+      - '.github/workflows/build-docker-images.yml'
+  workflow_dispatch:
+    inputs:
+      images:
+        description: 'Images to build (comma-separated: test-aarch64-a2, test-aarch64-a3)'
+        required: false
+        type: string
+        default: 'test-aarch64-a2'
+      push:
+        description: 'Push images to registry'
+        required: false
+        type: boolean
+        default: false
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+env:
+  REGISTRY: quay.io
+  QUAY_ORG: kerer
+  IMAGE_NAME: pytorch
+
+jobs:
+  matrix-prep:
+    runs-on: ubuntu-latest
+    outputs:
+      images: ${{ steps.select.outputs.images }}
+      do_push: ${{ steps.select.outputs.do_push }}
+    steps:
+      - name: Select images to build
+        id: select
+        run: |
+          ALL_IMAGES='["test-aarch64-a2","test-aarch64-a3"]'
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            INPUT="${{ github.event.inputs.images }}"
+            JSON=$(echo "${INPUT}" | sed 's/ *, */","/g' | sed 's/^/["/' | sed 's/$/"]/')
+            echo "images=${JSON}" >> "$GITHUB_OUTPUT"
+            echo "do_push=${{ github.event.inputs.push }}" >> "$GITHUB_OUTPUT"
+          else
+            echo "images=${ALL_IMAGES}" >> "$GITHUB_OUTPUT"
+            echo "do_push=true" >> "$GITHUB_OUTPUT"
+          fi
+
+  build:
+    needs: matrix-prep
+    environment: QUAY_USERNAME
+    runs-on: ubuntu-24.04-arm
+    strategy:
+      fail-fast: false
+      matrix:
+        image: ${{ fromJSON(needs.matrix-prep.outputs.images) }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Log in to Quay.io
+        if: needs.matrix-prep.outputs.do_push == 'true'
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ secrets.QUAY_USERNAME }}
+          password: ${{ secrets.QUAY_PASSWORD }}
+
+      - name: Build and push image
+        run: |
+          case "${{ matrix.image }}" in
+            test-aarch64-a2)
+              BASE_TAG="torch-npu-test-aarch64-cann-a2-py3.10-torch-nightly"
+              ;;
+            test-aarch64-a3)
+              BASE_TAG="torch-npu-test-aarch64-cann-a3-py3.10-torch-nightly"
+              ;;
+            *)
+              echo "ERROR: Unknown image: ${{ matrix.image }}"
+              exit 1
+              ;;
+          esac
+
+          chmod +x .ci/docker/docker_build.sh
+
+          echo "=== Building: ${BASE_TAG} ==="
+          bash .ci/docker/docker_build.sh "${BASE_TAG}"
+
+          # docker_build.sh appends a timestamp suffix (YYYYMMDDHHMM) to the
+          # tag. Docker treats the full string as the repository name with tag
+          # "latest", so we match on repository name.
+          BUILT_REPO="$(docker images --format '{{.Repository}}' \
+            | grep -E "^${BASE_TAG}-[0-9]{12}$" \
+            | sort -r | head -1)"
+
+          if [ -z "${BUILT_REPO}" ]; then
+            echo "ERROR: Could not find built image for ${BASE_TAG}"
+            docker images | grep "${BASE_TAG}"
+            exit 1
+          fi
+
+          TAG="${BUILT_REPO}:latest"
+
+          echo "=== Built: ${TAG} ==="
+          docker image inspect "${TAG}" --format '  Size: {{ .Size }}'
+
+          # Remote tag = the timestamped repository name (without :latest)
+          REMOTE_IMAGE="${REGISTRY}/${QUAY_ORG}/${IMAGE_NAME}:${BUILT_REPO}"
+
+          if [ "${{ needs.matrix-prep.outputs.do_push }}" = "true" ]; then
+            docker tag "${TAG}" "${REMOTE_IMAGE}"
+            docker push "${REMOTE_IMAGE}"
+            echo "=== Pushed: ${REMOTE_IMAGE} ==="
+          fi
+
+          mkdir -p /tmp/result
+          echo "${REMOTE_IMAGE}" > "/tmp/result/${BUILT_REPO}.txt"
+
+      - name: Upload result
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: result-${{ matrix.image }}
+          path: /tmp/result/*.txt
+          retention-days: 1
+
+  summary:
+    needs: [matrix-prep, 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 "" >> $GITHUB_STEP_SUMMARY
+          echo "**Registry:** \`${REGISTRY}/${QUAY_ORG}/${IMAGE_NAME}\`" >> $GITHUB_STEP_SUMMARY
diff --git a/.github/workflows/test-build-trigger.yml b/.github/workflows/test-build-trigger.yml
new file mode 100644
index 0000000000..2a13780151
--- /dev/null
+++ b/.github/workflows/test-build-trigger.yml
@@ -0,0 +1,29 @@
+name: Test Build Trigger
+
+on:
+  pull_request:
+    branches:
+      - master
+
+jobs:
+  resolve:
+    runs-on: ubuntu-latest
+    outputs:
+      sha: ${{ steps.resolve.outputs.sha }}
+    steps:
+      - name: Resolve upstream PyTorch main HEAD
+        id: resolve
+        run: |
+          SHA=$(git ls-remote https://github.com/pytorch/pytorch.git refs/heads/main | awk '{print $1}')
+          echo "Upstream PyTorch main HEAD: ${SHA}"
+          echo "sha=${SHA}" >> "$GITHUB_OUTPUT"
+
+  build:
+    needs: resolve
+    uses: ./.github/workflows/_build.yml
+    with:
+      upstream_repo: pytorch/pytorch
+      upstream_sha: ${{ needs.resolve.outputs.sha }}
+      upstream_fork_repo: ''
+      downstream_repo: ${{ github.event.pull_request.head.repo.full_name }}
+      downstream_ref: ${{ github.head_ref }}
diff --git a/ci/build.sh b/ci/build.sh
index 5a855130fe..54422c64aa 100644
--- a/ci/build.sh
+++ b/ci/build.sh
@@ -43,6 +43,11 @@ function parse_script_args() {
             args_num=$((args_num-1))
             shift
             ;;
+        --update_aoti_c_shim)
+            export UPDATE_AOTI_C_SHIM="--update_aoti_c_shim"
+            args_num=$((args_num-1))
+            shift
+            ;;
         --enable_lto)
             export ENABLE_LTO=TRUE
             args_num=$((args_num-1))
diff --git a/generate_code.sh b/generate_code.sh
index 946d44d49e..0c19b8cb75 100644
--- a/generate_code.sh
+++ b/generate_code.sh
@@ -8,6 +8,7 @@ cd $CDIR
 
 python_execute="$1"
 pytorch_version="$2"
+update_aoti_c_shim="${3:-${UPDATE_AOTI_C_SHIM:-}}"
 
 IFS='.' read -ra version_parts <<< "$pytorch_version"
 
@@ -57,12 +58,22 @@ testing_source_yaml="$CDIR/test/ops_unsupport_list.yaml"
 
 op_plugin_functions_yaml_path="$op_plugin_config_path/npu_native_functions.yaml"
 
-${python_execute} -m torchnpugen.gen_backend_stubs  \
-  --output_dir="torch_npu/csrc/aten" \
-  --source_yaml="$source_yaml" \
-  --impl_path="$CDIR/torch_npu/csrc/aten" \
-  --op_plugin_impl_path="$CDIR/third_party/op-plugin/op_plugin/ops" \
-  --op_plugin_yaml_path="$op_plugin_config_path/op_plugin_functions.yaml"
+if [ "${update_aoti_c_shim}" = "--update_aoti_c_shim" ]; then
+  ${python_execute} -m torchnpugen.gen_backend_stubs  \
+    --output_dir="torch_npu/csrc/aten" \
+    --source_yaml="$source_yaml" \
+    --impl_path="$CDIR/torch_npu/csrc/aten" \
+    --op_plugin_impl_path="$CDIR/third_party/op-plugin/op_plugin/ops" \
+    --op_plugin_yaml_path="$op_plugin_config_path/op_plugin_functions.yaml" \
+    --update_aoti_c_shim
+else
+  ${python_execute} -m torchnpugen.gen_backend_stubs  \
+    --output_dir="torch_npu/csrc/aten" \
+    --source_yaml="$source_yaml" \
+    --impl_path="$CDIR/torch_npu/csrc/aten" \
+    --op_plugin_impl_path="$CDIR/third_party/op-plugin/op_plugin/ops" \
+    --op_plugin_yaml_path="$op_plugin_config_path/op_plugin_functions.yaml"
+fi
 
 ${python_execute} -m torchnpugen.autograd.gen_autograd \
   --out_dir="$CDIR/torch_npu/csrc/aten" \
diff --git a/torch_npu/_inductor/ascend_npu_ir/ascend_npu_ir/npu/inductor_patch/lowering.py b/torch_npu/_inductor/ascend_npu_ir/ascend_npu_ir/npu/inductor_patch/lowering.py
index 2842ae9ba2..c9f0c1fa9c 100644
--- a/torch_npu/_inductor/ascend_npu_ir/ascend_npu_ir/npu/inductor_patch/lowering.py
+++ b/torch_npu/_inductor/ascend_npu_ir/ascend_npu_ir/npu/inductor_patch/lowering.py
@@ -3068,9 +3068,6 @@ def is_aligned(x):
 make_fallback(aten.to_sparse)
 make_fallback(aten._to_sparse)
 
-# Needs dimname support
-make_fallback(aten.zeros.names)
-
 # 6) Pattern-matched
 make_fallback(
     aten._scaled_dot_product_efficient_attention.default,
diff --git a/torch_npu/csrc/aten/AutoCastOps.cpp b/torch_npu/csrc/aten/AutoCastOps.cpp
index 85c75bcb9c..8f9dd136ff 100644
--- a/torch_npu/csrc/aten/AutoCastOps.cpp
+++ b/torch_npu/csrc/aten/AutoCastOps.cpp
@@ -108,11 +108,8 @@ TORCH_LIBRARY_IMPL(aten, AutocastPrivateUse1, m) {
     // fp32_set_opt_dtype
     KERNEL_PRIVATEUSEONE(prod, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(prod, dim_int, fp32_set_opt_dtype)
-    KERNEL_PRIVATEUSEONE(prod, dim_Dimname, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(softmax, int, fp32_set_opt_dtype)
-    KERNEL_PRIVATEUSEONE(softmax, Dimname, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(log_softmax, int, fp32_set_opt_dtype)
-    KERNEL_PRIVATEUSEONE(log_softmax, Dimname, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(cumprod, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(cumprod, dimname, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(cumsum, fp32_set_opt_dtype)
@@ -124,7 +121,6 @@ TORCH_LIBRARY_IMPL(aten, AutocastPrivateUse1, m) {
     // when autocasting.
     KERNEL_PRIVATEUSEONE(sum, fp32_set_opt_dtype)
     KERNEL_PRIVATEUSEONE(sum, dim_IntList, fp32_set_opt_dtype)
-    KERNEL_PRIVATEUSEONE(sum, dim_DimnameList, fp32_set_opt_dtype)
     // fp32_append_dtype
     // The fp32_append_dtype wrapper overrides implicit promotion behavior.
     // norm does not implicitly promote, but be aware when adding new ops to this policy.
@@ -136,10 +132,6 @@ TORCH_LIBRARY_IMPL(aten, AutocastPrivateUse1, m) {
         at::Tensor (const at::Tensor &, const c10::optional<c10::Scalar>&, at::IntArrayRef, bool),
         at::Tensor (const at::Tensor &, const c10::optional<c10::Scalar>&, at::IntArrayRef, bool, at::ScalarType),
         fp32_append_dtype)
-    KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_PRIVATEUSEONE(ADD_NS(norm), "norm.names_ScalarOpt_dim",
-    at::Tensor (const at::Tensor &, const c10::optional<c10::Scalar>&, at::DimnameList, bool),
-        at::Tensor (const at::Tensor &, const c10::optional<c10::Scalar>&, at::DimnameList, bool, at::ScalarType),
-        fp32_append_dtype)
     // promote
     KERNEL_PRIVATEUSEONE(addcdiv, promote)
     KERNEL_PRIVATEUSEONE(addcmul, promote)
diff --git a/torch_npu/csrc/aten/VariableFallbackKernel.cpp b/torch_npu/csrc/aten/VariableFallbackKernel.cpp
index 25d96107e4..13f95c0a30 100644
--- a/torch_npu/csrc/aten/VariableFallbackKernel.cpp
+++ b/torch_npu/csrc/aten/VariableFallbackKernel.cpp
@@ -132,7 +132,7 @@ static void npuBasicAutogradNotImplementedFallbackImpl(
     // by putting it after the requires_grad checks.
     any_input_requires_grad = any_input_requires_grad && at::GradMode::is_enabled();
 
-    std::shared_ptr<WarnNotImplemented> grad_fn;
+    c10::intrusive_ptr<WarnNotImplemented> grad_fn;
     if (any_input_requires_grad) {
         // NB: It is standard to collect edges from all tensors
         // (see generated/VariableTypeEverything.cpp for examples)
@@ -144,9 +144,7 @@ static void npuBasicAutogradNotImplementedFallbackImpl(
             stack,
             stack_start,
             num_arguments);
-        grad_fn = std::shared_ptr<WarnNotImplemented>(
-            new WarnNotImplemented(op_name, all_tensors_on_stack.size()),
-            torch::autograd::deleteNode);
+        grad_fn = c10::make_intrusive<WarnNotImplemented>(op_name, all_tensors_on_stack.size());
         grad_fn->set_next_edges(torch::autograd::collect_next_edges(all_tensors_on_stack));
     }
 
diff --git a/torch_npu/csrc/aten/common/CopyKernel.cpp b/torch_npu/csrc/aten/common/CopyKernel.cpp
index 6f3e239fa4..8d47115180 100644
--- a/torch_npu/csrc/aten/common/CopyKernel.cpp
+++ b/torch_npu/csrc/aten/common/CopyKernel.cpp
@@ -439,12 +439,6 @@ at::Tensor& NPUNativeFunctions::copy_(at::Tensor& self, const at::Tensor& src, b
     if (src._is_zerotensor()) {
         return self.zero_();
     }
-    // save tensor dim name
-    c10::optional<at::DimnameList> names = src.opt_names();
-    if (names.has_value()) {
-        internal_set_names_inplace(self, names);
-    }
-
     if (torch_npu::utils::is_npu(self)) {
         if (torch_npu::utils::is_npu(src)) {
             copy_d2d(self, src, non_blocking);
diff --git a/torch_npu/csrc/aten/common/ResizeNpu.cpp b/torch_npu/csrc/aten/common/ResizeNpu.cpp
index a9bfc5e601..a27c98772b 100644
--- a/torch_npu/csrc/aten/common/ResizeNpu.cpp
+++ b/torch_npu/csrc/aten/common/ResizeNpu.cpp
@@ -1,5 +1,4 @@
 #include <ATen/ATen.h>
-#include <ATen/NamedTensorUtils.h>
 
 #include "torch_npu/csrc/framework/FormatHelper.h"
 #include "torch_npu/csrc/aten/NPUNativeFunctions.h"
@@ -8,40 +7,11 @@
 namespace at_npu {
 namespace native {
 
-inline const at::Tensor& resize_named_tensor_(
-    const at::Tensor& self,
-    c10::IntArrayRef size,
-    c10::optional<c10::MemoryFormat> format)
-{
-    TORCH_INTERNAL_ASSERT(self.has_names());
-    TORCH_CHECK(
-        self.sizes() == size,
-        "Cannot resize named tensor with resize_ or resize_as_ (tried to resize "
-        "Tensor",
-        self.names(),
-        " with size ",
-        self.sizes(),
-        " to ",
-        size,
-        "). This may be caused by passing a named tensor ",
-        "as an `out=` argument; please ensure that the sizes are the same. ",
-        OPS_ERROR(ErrCode::VALUE));
-    TORCH_CHECK(
-        !format.has_value(),
-        "Unsupported memory format for named tensor resize ",
-        format.value(),
-        OPS_ERROR(ErrCode::NOT_SUPPORT));
-    return self;
-}
-
 const at::Tensor& NPUNativeFunctions::resize_(
     const at::Tensor& self,
     c10::IntArrayRef size,
     c10::optional<c10::MemoryFormat> format)
 {
-    if (self.has_names()) {
-        return resize_named_tensor_(self, size, format);
-    }
     // because of resize _impl_npu_ only support at base format, so
     // no need to reflush NpuStorageDesc here.
     auto ks = self.key_set();
@@ -76,7 +46,6 @@ const at::Tensor& NPUNativeFunctions::resize_as_(
     }
 
     const at::Tensor& result = self.resize_(the_template.sizes());
-    at::namedinference::propagate_names(result, the_template);
     return result;
 }
 
diff --git a/torch_npu/csrc/aten/common/TensorFactories.cpp b/torch_npu/csrc/aten/common/TensorFactories.cpp
index 1a327c97e1..558f5789da 100644
--- a/torch_npu/csrc/aten/common/TensorFactories.cpp
+++ b/torch_npu/csrc/aten/common/TensorFactories.cpp
@@ -11,7 +11,6 @@
 #include <string>
 
 #include <ATen/ATen.h>
-#include <ATen/NamedTensorUtils.h>
 #include <c10/util/irange.h>
 #include <ATen/record_function.h>
 #include <c10/core/impl/COW.h>
@@ -325,10 +324,6 @@ at::Tensor empty_like_npu(
         }
     }
 
-    if (self.opt_names()) {
-        at::namedinference::propagate_names(result, self.names());
-    }
-
     return result;
 }
 
@@ -427,43 +422,6 @@ at::Tensor NPUNativeFunctions::unsafe_empty_with_format(
     return NPUNativeFunctions::empty_with_format(size, dtype_opt, layout_opt, device_opt, pin_memory_opt, dst_format, c10::nullopt);
 }
 
-at::Tensor NPUNativeFunctions::empty_with_format(
-    c10::IntArrayRef size,
-    c10::optional<at::DimnameList> names,
-    c10::optional<at::ScalarType> dtype_opt,
-    c10::optional<c10::Layout> layout_opt,
-    c10::optional<c10::Device> device_opt,
-    c10::optional<bool> pin_memory_opt,
-    int64_t dst_format)
-{
-    torch_npu::utils::torch_check_npu(c10::device_or_default(device_opt));
-    caffe2::TypeMeta dtype = c10::scalarTypeToTypeMeta(dtype_or_default(dtype_opt));
-    c10::TensorOptions options = c10::TensorOptions().dtype(dtype_opt)
-                                        .device(device_opt)
-                                        .layout(layout_opt)
-                                        .pinned_memory(pin_memory_opt);
-    at::Tensor result = OpPreparation::ApplyTensorWithFormat(size, options, dst_format);
-    if (names.has_value()) {
-        internal_set_names_inplace(result, names);
-    }
-
-    return result;
-}
-
-at::Tensor empty_with_format_name_npu(
-    c10::IntArrayRef size,
-    c10::optional<at::DimnameList> names,
-    const c10::TensorOptions &options,
-    int64_t dst_format)
-{
-    at::Tensor result = OpPreparation::ApplyTensorWithFormat(size, options, dst_format);
-    if (names.has_value()) {
-        internal_set_names_inplace(result, names);
-    }
-
-    return result;
-}
-
 at::Tensor NPUNativeFunctions::empty_strided(
     c10::IntArrayRef size,
     c10::IntArrayRef stride,
diff --git a/torch_npu/csrc/aten/common/TensorShape.cpp b/torch_npu/csrc/aten/common/TensorShape.cpp
index 26053b37d0..f15cb4abf4 100644
--- a/torch_npu/csrc/aten/common/TensorShape.cpp
+++ b/torch_npu/csrc/aten/common/TensorShape.cpp
@@ -1,7 +1,6 @@
 #include <ATen/ATen.h>
 #include <ATen/ExpandUtils.h>
 #include <ATen/InferSize.h>
-#include <ATen/NamedTensorUtils.h>
 #include <ATen/NativeFunctions.h>
 #include <torch/library.h>
 #include <ATen/native/SparseTensorUtils.h>
@@ -105,7 +104,6 @@ at::Tensor alias_with_sizes_and_strides_npu(
         self_tmp_->set_storage_offset(self.storage_offset());
         self_tmp_->set_sizes_and_strides(sizes, strides);
     }
-    at::namedinference::propagate_names(self_, self);
     return self_;
 }
 
@@ -193,8 +191,6 @@ at::Tensor NPUNativeFunctions::squeeze(const at::Tensor& self)
 {
     auto g = inferSqueezeGeometry(self);
     at::Tensor result = self.as_strided(std::get<0>(g), std::get<1>(g));
-    auto maybe_outnames = at::namedinference::compute_squeeze_outnames(self);
-    at::namedinference::propagate_names_if_nonempty(result, maybe_outnames);
     return result;
 }
 
@@ -207,7 +203,6 @@ at::Tensor NPUNativeFunctions::squeeze(const at::Tensor& self, int64_t dim)
     }
     auto g = inferSqueezeGeometry(self, dim);
     auto result = self.as_strided(std::get<0>(g), std::get<1>(g));
-    at::namedinference::propagate_names_except(result, self, {dim});
     return result;
 }
 
diff --git a/torch_npu/csrc/aten/npu_native_functions.yaml b/torch_npu/csrc/aten/npu_native_functions.yaml
index 839f98d15c..8b72b2c1a8 100644
--- a/torch_npu/csrc/aten/npu_native_functions.yaml
+++ b/torch_npu/csrc/aten/npu_native_functions.yaml
@@ -21,9 +21,7 @@ supported:
   - empty_like
   - empty_strided
   - empty_with_format
-  - empty_with_format.names
   - full
-  - full.names
   - full.out
   - hamming_window
   - hamming_window.periodic
@@ -76,9 +74,6 @@ custom:
   - func: unsafe_empty_with_format(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, int acl_format=2, bool keep_format=False) -> Tensor
     dispatch:
       CompositeExplicitAutograd: empty_with_format
-  - func: empty_with_format.names(int[] size, Dimname[]? names, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, int acl_format=2) -> Tensor
-    dispatch:
-      CompositeExplicitAutograd: empty_with_format
   - func: copy_memory_(Tensor(a!) self, Tensor src, bool non_blocking=False) -> Tensor(a!)
     device_check: NoCheck
   - func: get_storage_size(Tensor self) -> int
diff --git a/torch_npu/csrc/aten/ops/FullKernelNpu.cpp b/torch_npu/csrc/aten/ops/FullKernelNpu.cpp
index 43682288ba..01cbc40417 100644
--- a/torch_npu/csrc/aten/ops/FullKernelNpu.cpp
+++ b/torch_npu/csrc/aten/ops/FullKernelNpu.cpp
@@ -1,5 +1,3 @@
-#include <ATen/NamedTensorUtils.h>
-
 #include "torch_npu/csrc/framework/utils/OpAdapter.h"
 #include "torch_npu/csrc/framework/utils/CalcuOpUtil.h"
 #include "torch_npu/csrc/aten/NPUNativeFunctions.h"
@@ -18,35 +16,5 @@ at::Tensor& NPUNativeFunctions::full_out(at::IntArrayRef size, const at::Scalar&
     return out;
 }
 
-at::Tensor NPUNativeFunctions::full(
-    at::IntArrayRef size,
-    const at::Scalar& fill_value,
-    c10::optional<at::DimnameList> names,
-    c10::optional<at::ScalarType> dtype_opt,
-    c10::optional<at::Layout> layout_opt,
-    c10::optional<at::Device> device_opt,
-    c10::optional<bool> pin_memory_opt)
-{
-    c10::TensorOptions option = c10::TensorOptions().dtype(dtype_opt)
-                                                    .device(device_opt)
-                                                    .layout(layout_opt)
-                                                    .pinned_memory(pin_memory_opt);
-    at::Tensor result = OpPreparation::ApplyTensorWithSizes(size, option);
-
-    if (!dtype_opt.has_value()) {
-        if (fill_value.isBoolean()) {
-            option = option.dtype(at::kBool);
-        } else if (fill_value.isIntegral(false)) {
-            option = option.dtype(at::kLong);
-        } else {
-            option = option.dtype(c10::get_default_dtype());
-        }
-    }
-
-    auto maybe_name = names.value_or(at::ArrayRef<at::Dimname>{});
-    at::namedinference::propagate_names_if_nonempty(result, maybe_name);
-    return result.fill_(fill_value);
-}
-
 }
 }
diff --git a/torch_npu/csrc/aten/ops/op_api/CloneKernelOpApi.cpp b/torch_npu/csrc/aten/ops/op_api/CloneKernelOpApi.cpp
index d73815eb1f..50a3ffeaeb 100644
--- a/torch_npu/csrc/aten/ops/op_api/CloneKernelOpApi.cpp
+++ b/torch_npu/csrc/aten/ops/op_api/CloneKernelOpApi.cpp
@@ -39,7 +39,6 @@ at::Tensor NPUNativeOpApiFunctions::clone(const at::Tensor &src, c10::optional<c
     }
 
     self.copy_(src);
-    at::namedinference::propagate_names(self, src);
     return self;
 }
 
diff --git a/torch_npu/csrc/aten/ops/op_api/CopyKernelOpApi.cpp b/torch_npu/csrc/aten/ops/op_api/CopyKernelOpApi.cpp
index 39604c6233..31a612b063 100644
--- a/torch_npu/csrc/aten/ops/op_api/CopyKernelOpApi.cpp
+++ b/torch_npu/csrc/aten/ops/op_api/CopyKernelOpApi.cpp
@@ -176,8 +176,6 @@ at::Tensor& NPUNativeOpApiFunctions::copy_(at::Tensor& self, const at::Tensor& s
         return self.zero_();
     }
 
-    auto maybe_outnames = at::namedinference::compute_broadcast_outnames(self, src);
-
     if (torch_npu::utils::is_npu(self)) {
         if (torch_npu::utils::is_npu(src)) {
             copy_d2d_baseformat_opapi(self, src, non_blocking);
@@ -200,7 +198,6 @@ at::Tensor& NPUNativeOpApiFunctions::copy_(at::Tensor& self, const at::Tensor& s
             }
         }
     }
-    at::namedinference::propagate_names_if_nonempty(self, maybe_outnames);
     return self;
 }
 
diff --git a/torch_npu/csrc/distributed/reducer.hpp b/torch_npu/csrc/distributed/reducer.hpp
index 904608e392..35ce1d2886 100644
--- a/torch_npu/csrc/distributed/reducer.hpp
+++ b/torch_npu/csrc/distributed/reducer.hpp
@@ -249,11 +249,11 @@ class Reducer {
     // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
     std::vector<bool> expect_sparse_gradients_;
 
-    std::vector<std::shared_ptr<torch::autograd::Node>>
+    std::vector<c10::intrusive_ptr<torch::autograd::Node>>
         grad_accumulators_; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
     // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
     std::unordered_map<torch::autograd::Node*, size_t> gradAccToVariableMap_;
-    std::vector<std::pair<uintptr_t, std::shared_ptr<torch::autograd::Node>>>
+    std::vector<std::pair<uintptr_t, c10::intrusive_ptr<torch::autograd::Node>>>
         hooks_; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
 
     // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
diff --git a/torch_npu/csrc/framework/autograd/VariableTypeManual.cpp b/torch_npu/csrc/framework/autograd/VariableTypeManual.cpp
index 7ef8237d0f..f463f44524 100644
--- a/torch_npu/csrc/framework/autograd/VariableTypeManual.cpp
+++ b/torch_npu/csrc/framework/autograd/VariableTypeManual.cpp
@@ -98,9 +98,9 @@ namespace {
 Tensor _fw_primal(c10::DispatchKeySet ks, const Tensor& self, int64_t level)
 {
     auto& self_ = unpack(self, "self", 0);
-    std::shared_ptr<Identity> grad_fn;
+    c10::intrusive_ptr<Identity> grad_fn;
     if (compute_requires_grad(self)) {
-        grad_fn = std::make_shared<Identity>();
+        grad_fn = c10::make_intrusive<Identity>();
         grad_fn->set_next_edges(collect_next_edges(self));
     }
 
diff --git a/torch_npu/csrc/framework/contiguous/ContiguousOpt.cpp b/torch_npu/csrc/framework/contiguous/ContiguousOpt.cpp
index 9ca6676477..1b45b1ab45 100644
--- a/torch_npu/csrc/framework/contiguous/ContiguousOpt.cpp
+++ b/torch_npu/csrc/framework/contiguous/ContiguousOpt.cpp
@@ -227,7 +227,6 @@ at::Tensor TransContiguous::view_tensor(const at::Tensor& self,
     auto* self_tmp_ = self_.unsafeGetTensorImpl();
     self_tmp_->set_storage_offset(offset);
     self_tmp_->set_sizes_and_strides(sizes, strides);
-    at::namedinference::propagate_names(self_, self);
     return self_;
 }
 
diff --git a/torch_npu/csrc/framework/contiguous/combined_opt.cpp b/torch_npu/csrc/framework/contiguous/combined_opt.cpp
index d95fd051db..2665f90d50 100644
--- a/torch_npu/csrc/framework/contiguous/combined_opt.cpp
+++ b/torch_npu/csrc/framework/contiguous/combined_opt.cpp
@@ -1,4 +1,3 @@
-#include <ATen/NamedTensorUtils.h>
 #include <ATen/quantized/QTensorImpl.h>
 
 #include <map>
diff --git a/torch_npu/csrc/framework/utils/CalcuOpUtil.h b/torch_npu/csrc/framework/utils/CalcuOpUtil.h
index 9f0cb38e10..a44fe6310a 100644
--- a/torch_npu/csrc/framework/utils/CalcuOpUtil.h
+++ b/torch_npu/csrc/framework/utils/CalcuOpUtil.h
@@ -6,7 +6,6 @@
 #include <vector>
 
 #include <ATen/ATen.h>
-#include <ATen/NamedTensorUtils.h>
 
 #include "torch_npu/csrc/core/npu/NPUException.h"
 #include "torch_npu/csrc/core/npu/DeviceUtils.h"
diff --git a/torch_npu/csrc/inductor/aoti_torch/generated/c_shim_npu.h b/torch_npu/csrc/inductor/aoti_torch/generated/c_shim_npu.h
index 104337efc4..c97c4861bf 100644
--- a/torch_npu/csrc/inductor/aoti_torch/generated/c_shim_npu.h
+++ b/torch_npu/csrc/inductor/aoti_torch/generated/c_shim_npu.h
@@ -24,7 +24,9 @@ AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__fft_c2c(AtenTensorHandle self,
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__fft_r2c(AtenTensorHandle self, const int64_t* dim, int64_t dim_len_, int64_t normalization, int32_t onesided, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__fused_moving_avg_obs_fq_helper_functional(AtenTensorHandle self, AtenTensorHandle observer_on, AtenTensorHandle fake_quant_on, AtenTensorHandle running_min, AtenTensorHandle running_max, AtenTensorHandle scale, AtenTensorHandle zero_point, double averaging_const, int64_t quant_min, int64_t quant_max, int64_t ch_axis, int32_t per_row_fake_quant, int32_t symmetric_quant, AtenTensorHandle* ret0, AtenTensorHandle* ret1, AtenTensorHandle* ret2, AtenTensorHandle* ret3, AtenTensorHandle* ret4, AtenTensorHandle* ret5);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__fused_rms_norm(AtenTensorHandle input, const int64_t* normalized_shape, int64_t normalized_shape_len_, AtenTensorHandle* weight, double* eps, AtenTensorHandle* ret0, AtenTensorHandle* ret1);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__grouped_mm(AtenTensorHandle self, AtenTensorHandle mat2, AtenTensorHandle* offs, AtenTensorHandle* bias, int32_t* out_dtype, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__pdist_forward(AtenTensorHandle self, double p, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__thnn_fused_lstm_cell(AtenTensorHandle input_gates, AtenTensorHandle hidden_gates, AtenTensorHandle cx, AtenTensorHandle* input_bias, AtenTensorHandle* hidden_bias, AtenTensorHandle* ret0, AtenTensorHandle* ret1, AtenTensorHandle* ret2);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu__to_sparse(AtenTensorHandle self, int32_t* layout, const int64_t** blocksize, int64_t blocksize_len_, int64_t* dense_dim, AtenTensorHandle* ret0);
@@ -59,6 +61,12 @@ AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_cumsum(AtenTensorHandle self, in
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_exponential(AtenTensorHandle self, double lambd, AtenGeneratorHandle* generator, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_fill__Scalar(AtenTensorHandle self, double value);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_grid_sampler_2d_backward(AtenTensorHandle grad_output, AtenTensorHandle input, AtenTensorHandle grid, int64_t interpolation_mode, int64_t padding_mode, int32_t align_corners, const int32_t* output_mask, int64_t output_mask_len_, AtenTensorHandle* ret0, AtenTensorHandle* ret1);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_13_0
+AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_grid_sampler_3d(AtenTensorHandle input, AtenTensorHandle grid, int64_t interpolation_mode, int64_t padding_mode, int32_t align_corners, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_13_0
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_13_0
+AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_grid_sampler_3d_backward(AtenTensorHandle grad_output, AtenTensorHandle input, AtenTensorHandle grid, int64_t interpolation_mode, int64_t padding_mode, int32_t align_corners, const int32_t* output_mask, int64_t output_mask_len_, AtenTensorHandle* ret0, AtenTensorHandle* ret1);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_13_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_hann_window(int64_t window_length, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_histc(AtenTensorHandle self, int64_t bins, double min, double max, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_index_Tensor(AtenTensorHandle self, const AtenTensorHandle** indices, int64_t indices_len_, AtenTensorHandle* ret0);
@@ -75,7 +83,9 @@ AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_max_pool3d_with_indices_backward
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_max_unpool2d(AtenTensorHandle self, AtenTensorHandle indices, const int64_t* output_size, int64_t output_size_len_, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_max_unpool3d(AtenTensorHandle self, AtenTensorHandle indices, const int64_t* output_size, int64_t output_size_len_, const int64_t* stride, int64_t stride_len_, const int64_t* padding, int64_t padding_len_, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_median(AtenTensorHandle self, AtenTensorHandle* ret0);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_11_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_mm_dtype_out(AtenTensorHandle out, AtenTensorHandle self, AtenTensorHandle mat2, int32_t out_dtype);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_11_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_mm_out(AtenTensorHandle out, AtenTensorHandle self, AtenTensorHandle mat2);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_mul_Scalar(AtenTensorHandle self, double other, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_mul_Tensor(AtenTensorHandle self, AtenTensorHandle other, AtenTensorHandle* ret0);
@@ -92,18 +102,30 @@ AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_pow_Tensor_Scalar(AtenTensorHand
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_pow_Tensor_Tensor(AtenTensorHandle self, AtenTensorHandle exponent, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_rand(const int64_t* size, int64_t size_len_, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_rand_generator(const int64_t* size, int64_t size_len_, AtenGeneratorHandle* generator, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_rand_like(AtenTensorHandle self, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_rand_like_generator(AtenTensorHandle self, AtenGeneratorHandle* generator, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint(int64_t high, const int64_t* size, int64_t size_len_, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint_generator(int64_t high, const int64_t* size, int64_t size_len_, AtenGeneratorHandle* generator, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint_low(int64_t low, int64_t high, const int64_t* size, int64_t size_len_, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint_low_out(AtenTensorHandle out, int64_t low, int64_t high, const int64_t* size, int64_t size_len_);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint_like(AtenTensorHandle self, int64_t high, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randint_like_low_dtype(AtenTensorHandle self, int64_t low, int64_t high, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randn(const int64_t* size, int64_t size_len_, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randn_generator(const int64_t* size, int64_t size_len_, AtenGeneratorHandle* generator, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randn_like(AtenTensorHandle self, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
+#if TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randn_like_generator(AtenTensorHandle self, AtenGeneratorHandle* generator, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, int32_t* memory_format, AtenTensorHandle* ret0);
+#endif // TORCH_FEATURE_VERSION >= TORCH_VERSION_2_12_0
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_randperm(int64_t n, int32_t* dtype, int32_t* layout, int32_t* device, int32_t device_index_, int32_t* pin_memory, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_replication_pad1d_backward(AtenTensorHandle grad_output, AtenTensorHandle self, const int64_t* padding, int64_t padding_len_, AtenTensorHandle* ret0);
 AOTI_TORCH_EXPORT AOTITorchError aoti_torch_npu_replication_pad2d_backward(AtenTensorHandle grad_output, AtenTensorHandle self, const int64_t* padding, int64_t padding_len_, AtenTensorHandle* ret0);
diff --git a/torchnpugen/autograd/templates/Functions.h b/torchnpugen/autograd/templates/Functions.h
index bb822779e6..fc1e0ed24b 100644
--- a/torchnpugen/autograd/templates/Functions.h
+++ b/torchnpugen/autograd/templates/Functions.h
@@ -27,7 +27,7 @@ using at::ScalarType;
 using c10::optional;
 using c10::fmap;
 
-inline std::vector<Tensor> unpack_list(at::ArrayRef<SavedVariable> xs, std::shared_ptr<Node> saved_for = nullptr)
+inline std::vector<Tensor> unpack_list(at::ArrayRef<SavedVariable> xs, c10::intrusive_ptr<Node> saved_for = nullptr)
 {
     // NB: we must explicitly do the conversion in the lambda, otherwise template
     // deduction will give a Tensor of Variable which is not convertible
@@ -36,7 +36,7 @@ inline std::vector<Tensor> unpack_list(at::ArrayRef<SavedVariable> xs, std::shar
     });
 }
 
-inline c10::List<c10::optional<Tensor>> unpack_opt_list(at::ArrayRef<SavedVariable> xs, std::shared_ptr<Node> saved_for = nullptr)
+inline c10::List<c10::optional<Tensor>> unpack_opt_list(at::ArrayRef<SavedVariable> xs, c10::intrusive_ptr<Node> saved_for = nullptr)
 {
     torch::List<c10::optional<Tensor>> result;
     result.reserve(xs.size());
diff --git a/torchnpugen/autograd/templates/VariableType.h b/torchnpugen/autograd/templates/VariableType.h
index ec552d867b..3c0e896dca 100644
--- a/torchnpugen/autograd/templates/VariableType.h
+++ b/torchnpugen/autograd/templates/VariableType.h
@@ -25,8 +25,6 @@ namespace at_npu { namespace autograd {
 using Variable = at::Tensor;
 using at::Context;
 using at::Device;
-using at::Dimname;
-using at::DimnameList;
 using at::Generator;
 using at::IntArrayRef;
 using at::MemoryFormat;
diff --git a/torchnpugen/templates/CustomRegisterSchema.cpp b/torchnpugen/templates/CustomRegisterSchema.cpp
index 15bd3fa4aa..e62224039f 100644
--- a/torchnpugen/templates/CustomRegisterSchema.cpp
+++ b/torchnpugen/templates/CustomRegisterSchema.cpp
@@ -3,7 +3,6 @@
 #include <ATen/DeviceGuard.h>
 #include <ATen/NativeFunctions.h>
 #include <ATen/MetaFunctions.h>
-#include <ATen/NamedTensorUtils.h>
 #include <ATen/Utils.h>
 #include <ATen/WrapDimUtils.h>
 #include <ATen/Dispatch.h>