chore: add VS Code devcontainer configuration.-MindSpeed-MM-AtomGit

A.devcontainer/README.md+65-0

@@ -0,0 +1,65 @@
	1	+# MindSpeed-MM Dev Container
	2	+
	3	+在 VS Code 中一键启动 MindSpeed-MM 开发环境。
	yaoyaoxu8月1日建议补充说明这个功能的背景与利弊 X xubin 8月1日评论： 8月1日评论：
	4	+
	5	+## 工作模式
	6	+
	7	+[devcontainer.json](./devcontainer.json) 默认基于预先构建好的镜像（`image` 字段），拉取即用，无需本地构建。
	8	+
	9	+> 配置文件中填写的 `image` 仅为示例，请根据实际需求选择所需镜像。镜像获取方式与更多信息请查看 [docker/OVERVIEW.zh.md](../docker/OVERVIEW.zh.md)。
	10	+
	11	+如需自行修改构建参数，请在 `devcontainer.json` 中：
	12	+
	13	+1. 注释掉 `image` 行
	14	+2. 取消注释 `build` 块与 `initializeCommand` 行
	15	+
	16	+即可切换为从 [../docker/Dockerfile](../docker/Dockerfile) 本地构建的模式。
	17	+
	18	+## 前置要求
	19	+
	20	+- [VS Code](https://code.visualstudio.com/) + [Dev Containers](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-containers) 扩展
	21	+- 本地安装 Docker
	22	+- 宿主机已安装昇腾 NPU 驱动，用于容器内访问 NPU
	23	+
	24	+## 使用方式
	25	+
	26	+1. 用 VS Code 打开本项目根目录
	yaoyaoxu8月1日是不是需要先ssh到服务器？ X xubin 8月1日评论： 8月1日评论：
	27	+2. 按 `F1` 打开命令面板，执行 Dev Containers: Reopen in Container
	28	+3. 首次启动会拉取预构建镜像（默认模式）或本地构建镜像（自建模式，约 20-40 分钟，取决于网络），完成后进入开发环境
	29	+
	30	+> 自建模式下，`initializeCommand` 会调用 [prepare-context.sh](./prepare-context.sh)，自动下载 Miniconda 并准备 Dockerfile 所需的构建上下文文件，无需手动操作。
	31	+
	32	+## 构建参数调整（自建模式）
	33	+
	34	+编辑 [devcontainer.json](./devcontainer.json) 的 `build.args` 即可调整构建参数：
	35	+如果参数只是一个示例，请根据实际需求填写。
	36	+
	37	+\| 参数 \| 默认值 \| 说明 \|
	38	+\| --- \| --- \| --- \|
	39	+\| `NPU_TYPE` \| `a3` \| NPU 型号：`910b` 或 `a3` \|
	40	+\| `OS` \| `openeuler24.03` \| 操作系统：`openeuler24.03` 或 `ubuntu22.04` \|
	41	+\| `BASE_IMAGE_VERSION` \| `9.0.0` \| CANN 基础镜像版本 \|
	42	+\| `TORCH_VERSION` \| `2.7.1` \| PyTorch 版本（在线安装） \|
	43	+\| `TORCH_NPU_VERSION` \| `2.7.1.post8` \| torch_npu 版本（在线安装） \|
	44	+\| `MINDSPEED_MM_BRANCH` \| `26.1.0` \| MindSpeed-MM 分支 \|
	45	+
	46	+## NPU 设备配置
	47	+
	48	+默认挂载 `/dev/davinci0` 与 `/dev/davinci1`。如需使用其他卡号，修改 `devcontainer.json` 中 `runArgs` 的 `--device=/dev/davinci2` 索引，或新增 `--device=/dev/davinci2` 等。
	49	+
	50	+宿主机无 NPU 时，注释掉 `runArgs` 与 `mounts` 即可作为纯代码编辑环境使用。
	51	+
	52	+## 工作区说明
	53	+
	54	+- 本地源码通过 bind mount 挂载到容器 `/workspace/MindSpeed-MM`，编辑实时生效
	55	+- 容器创建后自动执行 `pip install -e . --no-deps` 以可编辑模式安装项目，并配置 pre-commit
	56	+- 默认 Python 解释器：`/opt/conda/bin/python`（conda base 环境）
	57	+
	58	+## 常见问题排查
	59	+
	60	+- Miniconda 下载超时（自建模式）
	61	+ 手动下载 [Miniconda3-py311_26.1.1-1-Linux-${ARCH}.sh](https://repo.anaconda.com/miniconda/) 放到 `docker/downloads/` 目录后重新构建
	62	+
	63	+- decord 依赖下载失败（仅 ARM 架构需要 & 自建模式）
	64	+ decord 编译依赖较多且常因网络问题下载失败，可用浏览器手动下载后放入 `docker/decord_deps/` 目录，再重新执行 Reopen in Container。下载链接参考 [docker/common/download_decord_deps.sh](../docker/common/download_decord_deps.sh)。
	65	+ 完成后 `docker/decord_deps/` 目录结构应包含上述全部文件与目录。x86_64 架构无需下载这些依赖（decord 在镜像构建阶段一已通过 pip 安装）。

A.devcontainer/devcontainer.json+85-0

@@ -0,0 +1,85 @@
	1	+{
	2	+ "name": "MindSpeed-MM Dev",
	3	+ "image": "swr.cn-south-1.myhuaweicloud.com/ascendhub/mindspeed-mm:26.0.0-a3-openeuler24.03-py3.11-aarch64",
	4	+ // // ----- build 模式（从 Dockerfile 构建时取消注释，注释上方 image 行） -----
	5	+ // "build": {
	6	+ // "dockerfile": "../docker/Dockerfile",
	7	+ // "context": "../docker",
	8	+ // // 构建时使用 host 网络，使构建容器可通过 127.0.0.1:6268 访问宿主机代理
	9	+ // "options": ["--network=host"],
	10	+ // "args": {
	11	+ // // PyTorch / torch_npu 版本（在线安装）
	12	+ // "TORCH_VERSION": "2.7.1",
	13	+ // "TORCH_NPU_VERSION": "2.7.1.post8",
	14	+ // // MindSpeed-MM 分支（dev 镜像会克隆该分支；工作区挂载会覆盖克隆内容）
	15	+ // "MINDSPEED_MM_BRANCH": "26.1.0",
	16	+ // // 构建开发镜像（非 CI 多版本镜像）
	17	+ // "BUILD_CI": "false",
	18	+ // // CANN 基础镜像版本
	19	+ // "BASE_IMAGE": "swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11",
	20	+ // // 以下参数对应 prepare-context.sh 暂存到上下文中的文件/目录名
	21	+ // "MINICONDA_SH": "miniconda.sh",
	22	+ // "DECORD_SCRIPT": "install_decord_on_arm.sh",
	23	+ // "DECORD_DEPS_DIR": "decord_deps",
	24	+ // "TORCH_WHL_DIR": "torch_wheels"
	25	+ // }
	26	+ // },
	27	+ // // build 模式专用：构建前在宿主机执行，下载 Miniconda、暂存辅助脚本
	28	+ // "initializeCommand": "bash .devcontainer/prepare-context.sh",
	29	+ // 将本地源码挂载到镜像默认工作目录，覆盖构建时克隆的仓库，实现实时编辑
	30	+ "workspaceMount": "source=${localWorkspaceFolder},target=/workspace/MindSpeed-MM,type=bind,consistency=cached",
	31	+ "workspaceFolder": "/workspace/MindSpeed-MM",
	32	+ "remoteUser": "root",
	33	+ // 昇腾 NPU 设备访问。请按宿主机实际卡号调整 /dev/davinci0 索引。
	34	+ // 若宿主机无 NPU，注释掉 runArgs 与 mounts 即可作为纯代码编辑环境使用。
	35	+ "runArgs": [
	36	+ "--network=host",
	37	+ "--device=/dev/davinci0",
	38	+ "--device=/dev/davinci1",
	39	+ "--device=/dev/davinci_manager",
	40	+ "--device=/dev/devmm_svm",
	41	+ "--device=/dev/hisi_hdc",
	42	+ "--ipc=host"
	43	+ ],
	44	+ "mounts": [
	45	+ "source=/usr/local/dcmi,target=/usr/local/dcmi,type=bind",
	46	+ "source=/usr/local/bin/npu-smi,target=/usr/local/bin/npu-smi,type=bind",
	47	+ "source=/usr/local/Ascend/driver/lib64,target=/usr/local/Ascend/driver/lib64,type=bind",
	48	+ "source=/usr/local/Ascend/driver/version.info,target=/usr/local/Ascend/driver/version.info,type=bind",
	49	+ "source=/etc/ascend_install.info,target=/etc/ascend_install.info,type=bind"
	50	+ ],
	51	+ // 容器创建后：以可编辑模式安装项目（依赖已在镜像中预装），并配置 pre-commit
	52	+ "postCreateCommand": "pip install -e . --no-deps && pip install pre-commit && pre-commit install",
	53	+ "customizations": {
	54	+ "vscode": {
	55	+ "extensions": [
	56	+ // Python 核心
	57	+ "ms-python.python",
	58	+ "ms-python.vscode-pylance",
	59	+ "ms-python.debugpy",
	60	+ // Ruff（linting + formatting）
	61	+ "charliermarsh.ruff",
	62	+ // 通用
	63	+ "eamodio.gitlens",
	64	+ "yzhang.markdown-all-in-one",
	65	+ "tamasfe.even-better-toml",
	66	+ "redhat.vscode-yaml"
	67	+ ],
	68	+ "settings": {
	69	+ "python.defaultInterpreterPath": "/opt/conda/bin/python",
	70	+ "python.terminal.activateEnvironment": true,
	71	+ "python.analysis.typeCheckingMode": "basic",
	72	+ "python.analysis.autoImportCompletions": true,
	73	+ "editor.formatOnSave": true,
	74	+ "editor.rulers": [120],
	75	+ "files.trimTrailingWhitespace": true,
	76	+ "files.insertFinalNewline": true,
	77	+ "yaml.schemas": {},
	78	+ "search.exclude": {
	79	+ "*/kernel_meta_": true,
	80	+ "**/.cache": true
	81	+ }
	82	+ }
	83	+ }
	84	+ }
	85	+}

A.devcontainer/prepare-context.sh+88-0

@@ -0,0 +1,88 @@
	1	+#!/bin/bash
	2	+set -e
	3	+
	4	+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
	5	+PROJECT_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
	6	+DOCKER_DIR="${PROJECT_ROOT}/docker"
	7	+COMMON_DIR="${DOCKER_DIR}/common"
	8	+MODEL_INSTALL_DIR="${DOCKER_DIR}/scripts/model_install"
	9	+
	10	+# 可通过环境变量覆盖默认值
	11	+OS="${OS:-openeuler24.03}"
	12	+
	13	+# 根据 OS 推导 OS_FAMILY 与软件源配置脚本
	14	+case "$OS" in
	15	+ openeuler*) OS_FAMILY="openeuler"; REPO_SCRIPT="configure_yum_repo.sh" ;;
	16	+ ubuntu*) OS_FAMILY="ubuntu"; REPO_SCRIPT="configure_apt_repo.sh" ;;
	17	+ *) echo "ERROR: 不支持的操作系统: $OS（可选: openeuler24.03 / ubuntu22.04）"; exit 1 ;;
	18	+esac
	19	+
	20	+# 检测 CPU 架构
	21	+ARCH="$(uname -m)"
	22	+if [ "$ARCH" != "x86_64" ] && [ "$ARCH" != "aarch64" ]; then
	23	+ echo "ERROR: 不支持的 CPU 架构: $ARCH（仅支持 x86_64 / aarch64）"
	24	+ exit 1
	25	+fi
	26	+
	27	+echo "=========================================="
	28	+echo "准备 devcontainer 构建上下文"
	29	+echo " OS: ${OS}"
	30	+echo " OS_FAMILY: ${OS_FAMILY}"
	31	+echo " 架构: ${ARCH}"
	32	+echo " 上下文目录: ${DOCKER_DIR}"
	33	+echo "=========================================="
	34	+
	35	+cd "${DOCKER_DIR}"
	36	+
	37	+# 1. 下载并暂存 Miniconda 安装器（统一命名为 miniconda.sh，便于构建参数固定）
	38	+MINICONDA_FILE="Miniconda3-py311_26.1.1-1-Linux-${ARCH}.sh"
	39	+if [ ! -f "miniconda.sh" ]; then
	40	+ if [ -f "downloads/${MINICONDA_FILE}" ]; then
	41	+ echo ">>> 复用已下载的 Miniconda: downloads/${MINICONDA_FILE}"
	42	+ else
	43	+ echo ">>> 下载 Miniconda (${ARCH})..."
	44	+ bash "${COMMON_DIR}/download_miniconda.sh" "${DOCKER_DIR}/downloads" "${ARCH}"
	45	+ fi
	46	+ cp "downloads/${MINICONDA_FILE}" "miniconda.sh"
	47	+else
	48	+ echo ">>> Miniconda 已存在: miniconda.sh"
	49	+fi
	50	+
	51	+# 2. 暂存 common_functions.sh
	52	+cp "${COMMON_DIR}/common_functions.sh" "common_functions.sh"
	53	+
	54	+# 3. 暂存软件源配置脚本为 configure_repo.sh（Dockerfile 固定 COPY 该名称）
	55	+cp "${COMMON_DIR}/${REPO_SCRIPT}" "configure_repo.sh"
	56	+
	57	+# 4. 暂存模型环境安装脚本到 install_scripts/
	58	+mkdir -p "install_scripts"
	59	+for script in "${MODEL_INSTALL_DIR}"/install_*.sh; do
	60	+ [ -f "$script" ] && cp "$script" "install_scripts/"
	61	+done
	62	+
	63	+# 5. 暂存 decord 安装脚本（ARM 构建需要；x86 构建会被跳过，但文件必须存在）
	64	+cp "${COMMON_DIR}/install_decord_on_arm.sh" "install_decord_on_arm.sh"
	65	+
	66	+# 6. 创建 torch_wheels 占位目录（空目录 = 在线安装 PyTorch）
	67	+mkdir -p "torch_wheels"
	68	+touch "torch_wheels/.placeholder"
	69	+
	70	+# 7. 创建 decord_deps 目录
	71	+# - x86_64: decord 已在阶段一通过 pip 安装，仅创建占位目录满足 COPY 路径
	72	+# - aarch64: 下载 decord 源码编译依赖
	73	+if [ "$ARCH" = "aarch64" ]; then
	74	+ if [ ! -d "decord_deps" ] \|\| [ -z "$(ls -A decord_deps 2>/dev/null)" ]; then
	75	+ echo ">>> 下载 decord 编译依赖(ARM)..."
	76	+ rm -rf "decord_deps"
	77	+ bash "${COMMON_DIR}/download_decord_deps.sh" "${DOCKER_DIR}/decord_deps"
	78	+ else
	79	+ echo ">>> decord 依赖已存在: decord_deps"
	80	+ fi
	81	+else
	82	+ mkdir -p "decord_deps"
	83	+ touch "decord_deps/.placeholder"
	84	+fi
	85	+
	86	+echo "=========================================="
	87	+echo "构建上下文准备完成"
	88	+echo "=========================================="

M.gitignore+12-1

	@@ -158,4 +158,15 @@ mindspeed
158	/tests/st/run_logs/	158	/tests/st/run_logs/
159		159
160	fusion_result.json	160	fusion_result.json
161	-cache_dir/	161	+cache_dir/
		162	+
		163	+# devcontainer: Build context artifacts staged to docker/ via .devcontainer/prepare-context.sh
		164	+/docker/miniconda.sh
		165	+/docker/Miniconda3-*.sh
		166	+/docker/common_functions.sh
		167	+/docker/configure_repo.sh
		168	+/docker/install_decord_on_arm.sh
		169	+/docker/install_scripts/
		170	+/docker/torch_wheels/
		171	+/docker/decord_deps/
		172	+/docker/downloads/

M.pre-commit-config.yaml+1-0

	@@ -19,6 +19,7 @@ repos:
19	- id: check-merge-conflict	19	- id: check-merge-conflict
20	- id: detect-private-key	20	- id: detect-private-key
21	- id: check-json	21	- id: check-json
		22	+ exclude: \.devcontainer/devcontainer\.json$
22		23
23	# -------------------------- Python 核心检查 --------------------------	24	# -------------------------- Python 核心检查 --------------------------
24	# codespell	25	# codespell

Mdocker/Dockerfile+17-29

	@@ -17,20 +17,15 @@
17	# then produce the final image.	17	# then produce the final image.
18	# ============================================	18	# ============================================
19		19
20	-# --- Base image arguments ---	20	+# BASE_IMAGE: Full base image URL; when empty, falls back to the hardcoded default below.
21	-# OS: Operating system type, e.g. openeuler24.03 or ubuntu22.04	21	+# NOTE: Do not use nested ${VAR} inside ${BASE_IMAGE:-default} — Docker's :- parser
22	-ARG OS=openeuler24.03	22	+# stops at the first '}' and produces a malformed image name.
23	-# BASE_IMAGE_VERSION: CANN base image version tag
24	-ARG BASE_IMAGE_VERSION=9.0.0-beta.2
25	-# NPU_TYPE: Ascend NPU model (910b / a3 / etc.)
26	-ARG NPU_TYPE=910b
27	-# BASE_IMAGE: Full base image URL; the default splicing rule is used when empty
28	ARG BASE_IMAGE=""	23	ARG BASE_IMAGE=""
29		24
30	# Stage 1 (base): Build the base runtime environment.	25	# Stage 1 (base): Build the base runtime environment.
31	# Pulls the CANN base image from the Huawei Cloud Ascend registry by default;	26	# Pulls the CANN base image from the Huawei Cloud Ascend registry by default;
32	# comes with CANN and Python 3.11 pre-installed.	27	# comes with CANN and Python 3.11 pre-installed.
33	-FROM ${BASE_IMAGE:-swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:${BASE_IMAGE_VERSION}-${NPU_TYPE}-${OS}-py3.11} AS base	28	+FROM ${BASE_IMAGE:-swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11} AS base
34		29
35	# Switch to root (the CANN base image defaults to a non-root user)	30	# Switch to root (the CANN base image defaults to a non-root user)
36	USER root	31	USER root
	@@ -40,11 +35,9 @@ USER root
40	ARG MINICONDA_SH	35	ARG MINICONDA_SH
41	# TORCH_VERSION / TORCH_NPU_VERSION: PyTorch and torch_npu versions	36	# TORCH_VERSION / TORCH_NPU_VERSION: PyTorch and torch_npu versions
42	ARG TORCH_VERSION=2.7.1	37	ARG TORCH_VERSION=2.7.1
43	-ARG TORCH_NPU_VERSION=2.7.1	38	+ARG TORCH_NPU_VERSION=2.7.1.post8
44	# TORCH_WHL_DIR: Directory containing local PyTorch .whl files (optional, for offline install)	39	# TORCH_WHL_DIR: Directory containing local PyTorch .whl files (optional, for offline install)
45	ARG TORCH_WHL_DIR=torch_wheels	40	ARG TORCH_WHL_DIR=torch_wheels
46	-# OS_FAMILY: Operating system family (openeuler / ubuntu), used to select the package manager
47	-ARG OS_FAMILY
48		41
49	# Detect CPU architecture and print info (x86_64 / aarch64)	42	# Detect CPU architecture and print info (x86_64 / aarch64)
50	RUN ARCH=$(uname -m) && \	43	RUN ARCH=$(uname -m) && \
	@@ -60,18 +53,6 @@ RUN ARCH=$(uname -m) && \
60	fi && \	53	fi && \
61	echo "=========================================="	54	echo "=========================================="
62		55
63	-# Configure DNS resolution (ensures the container can access the public internet).
64	-# Uses: domestic DNS (114) -> Google DNS (8.8.8.8) -> Alibaba DNS (223.5.5.5)
65	-RUN if [ -f /etc/resolv.conf ]; then \
66	- echo "Existing resolv.conf:" && cat /etc/resolv.conf; \
67	- else \
68	- echo "No resolv.conf found, creating one..."; \
69	- fi && \
70	- echo "nameserver 114.114.114.114" > /etc/resolv.conf && \
71	- echo "nameserver 8.8.8.8" >> /etc/resolv.conf && \
72	- echo "nameserver 223.5.5.5" >> /etc/resolv.conf && \
73	- echo "Updated resolv.conf:" && cat /etc/resolv.conf
74	-
75	# Configure system package repositories (via external script for yum/apt sources)	56	# Configure system package repositories (via external script for yum/apt sources)
76	COPY configure_repo.sh /tmp/configure_repo.sh	57	COPY configure_repo.sh /tmp/configure_repo.sh
77	RUN chmod +x /tmp/configure_repo.sh && \	58	RUN chmod +x /tmp/configure_repo.sh && \
	@@ -82,8 +63,16 @@ RUN chmod +x /tmp/configure_repo.sh && \
82	# - Build toolchain: gcc/g++, make, cmake (for compiling C++ extensions like decord)	63	# - Build toolchain: gcc/g++, make, cmake (for compiling C++ extensions like decord)
83	# - Network tools: git, wget, curl	64	# - Network tools: git, wget, curl
84	# - Media processing: ffmpeg and related dev libraries (for video decoding)	65	# - Media processing: ffmpeg and related dev libraries (for video decoding)
85	-RUN echo "Installing system dependencies..." && \	66	+# OS family is detected from /etc/os-release (ID field), so no build arg is needed.
86	- if [ "$OS_FAMILY" = "openeuler" ]; then \	67	+RUN . /etc/os-release && \
		68	+ DETECTED_OS_FAMILY="" && \
		69	+ ID_LOWER=$(echo "$ID" \| tr '[:upper:]' '[:lower:]') && \
		70	+ case "$ID_LOWER" in \
		71	+ openeuler) DETECTED_OS_FAMILY="openeuler" ;; \
		72	+ ubuntu) DETECTED_OS_FAMILY="ubuntu" ;; \
		73	+ esac && \
		74	+ echo "Installing system dependencies (ID=$ID, OS_FAMILY=$DETECTED_OS_FAMILY)..." && \
		75	+ if [ "$DETECTED_OS_FAMILY" = "openeuler" ]; then \
87	yum install -y \	76	yum install -y \
88	git \	77	git \
89	iproute \	78	iproute \
	@@ -96,7 +85,7 @@ RUN echo "Installing system dependencies..." && \
96	cmake \	85	cmake \
97	ffmpeg-devel \	86	ffmpeg-devel \
98	&& yum clean all; \	87	&& yum clean all; \
99	- elif [ "$OS_FAMILY" = "ubuntu" ]; then \	88	+ elif [ "$DETECTED_OS_FAMILY" = "ubuntu" ]; then \
100	apt-get install -y \	89	apt-get install -y \
101	git \	90	git \
102	iproute2 \	91	iproute2 \
	@@ -114,8 +103,7 @@ RUN echo "Installing system dependencies..." && \
114	libswscale-dev \	103	libswscale-dev \
115	&& apt-get clean; \	104	&& apt-get clean; \
116	else \	105	else \
117	- echo "ERROR: Unsupported OS: $OS_FAMILY"; \	106	+ echo "ERROR: Unsupported OS ID: $ID" && exit 1; \
118	- exit 1; \
119	fi	107	fi
120		108
121	# Install Miniconda (for creating and managing Python virtual environments).	109	# Install Miniconda (for creating and managing Python virtual environments).

Mdocker/OVERVIEW.md+19-27

	@@ -66,7 +66,7 @@ Taking `v26.1.0-cann9.0.0-torch_npu2.7.1-a3-openeuler24.03-py3.11-aarch64` as an
66	\| Python version \| `py3.11` \| Python 3.11 \|	66	\| Python version \| `py3.11` \| Python 3.11 \|
67	\| architecture \| `aarch64` \| ARM64 architecture \|	67	\| architecture \| `aarch64` \| ARM64 architecture \|
68		68
69	-> The CANN version and TorchNPU version are derived from the base image and the `--torch-npu-version` parameter at build time.	69	+> The CANN version, NPU type, OS, and Python version are all derived from the base image tag. The TorchNPU version comes from the `--torch-npu-version` parameter at build time.
70		70
71	Dockerfile naming follows the template: `Dockerfile[.{chip info}.{OS}.{other fields}]`	71	Dockerfile naming follows the template: `Dockerfile[.{chip info}.{OS}.{other fields}]`
72		72
	@@ -181,12 +181,11 @@ The image includes the following pre-configured environments:
181		181
182	### Build Script Parameter Description	182	### Build Script Parameter Description
183		183
184	-The build script `build.sh` supports multiple parameter configurations. The following default values are examples only; please adjust according to actual needs. The default value `9.0.0` for `--base-image-version` is an example only. It is strongly recommended to explicitly specify the base image matching your target environment via `--base-image` or `--base-image-version`.	184	+The build script `build.sh` supports multiple parameter configurations. The CANN version, NPU type, OS, and Python version are only obtained by parsing the `--base-image` tag — they cannot be specified manually.
185		185
186	-\| Parameter \| Description \| Default (Example) \|	186	+\| Parameter \| Description \| Default \|
187	\| ------ \| ------ \| ------------ \|	187	\| ------ \| ------ \| ------------ \|
188	-\| `-t, --npu-type` \| NPU type: A3 or 910B (auto-detected from `--base-image` if not specified) \| None (required) \|	188	+\| `--base-image` \| Required. Full base image name. CANN version, NPU type, OS, and Python version are auto-detected from the image tag \| None (required) \|
189	-\| `-o, --os` \| Operating system: openeuler24.03 or ubuntu22.04 (auto-detected from `--base-image` if not specified) \| openeuler24.03 \|
190	\| `-v, --version` \| MindSpeed MM version identifier, also used as Git branch name and script directory selection basis \| 26.1.0 \|	189	\| `-v, --version` \| MindSpeed MM version identifier, also used as Git branch name and script directory selection basis \| 26.1.0 \|
191	\| `-m, --miniconda` \| Miniconda installer path (auto-downloaded if not specified) \| None (auto-download) \|	190	\| `-m, --miniconda` \| Miniconda installer path (auto-downloaded if not specified) \| None (auto-download) \|
192	\| `-d, --decord-deps` \| decord dependencies directory path (auto-downloaded for ARM if not specified) \| None (auto-download) \|	191	\| `-d, --decord-deps` \| decord dependencies directory path (auto-downloaded for ARM if not specified) \| None (auto-download) \|
	@@ -200,8 +199,6 @@ The build script `build.sh` supports multiple parameter configurations. The foll
200	\| `--torch-npu-whl` \| torch-npu.whl file path (offline installation) \| None \|	199	\| `--torch-npu-whl` \| torch-npu.whl file path (offline installation) \| None \|
201	\| `--torchvision-whl` \| torchvision.whl file path (optional, offline installation) \| None \|	200	\| `--torchvision-whl` \| torchvision.whl file path (optional, offline installation) \| None \|
202	\| `--torchaudio-whl` \| torchaudio.whl file path (optional, offline installation) \| None \|	201	\| `--torchaudio-whl` \| torchaudio.whl file path (optional, offline installation) \| None \|
203	-\| `--base-image-version` \| Base image CANN version (example value; recommended to specify explicitly) \| 9.0.0 \|
204	-\| `--base-image` \| Complete base image name (higher priority than `--base-image-version`; recommended) \| None \|
205	\| `--build-ci` \| Build CI image with multi-version conda environments (skip verl + MindSpeed-MM clone) \| None \|	202	\| `--build-ci` \| Build CI image with multi-version conda environments (skip verl + MindSpeed-MM clone) \| None \|
206	\| `--cleanup-on-fail` \| Clean up dangling images/containers on build failure \| None \|	203	\| `--cleanup-on-fail` \| Clean up dangling images/containers on build failure \| None \|
207		204
	@@ -210,28 +207,22 @@ The build script `build.sh` supports multiple parameter configurations. The foll
210	```bash	207	```bash
211	cd docker	208	cd docker
212		209
213	-# Build A3 + openEuler image (default)	210	+# Build A3 + openEuler image
214	-bash build.sh -t A3	211	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11
215	-
216	-# Build A2 + openEuler image
217	-bash build.sh -t 910B
218		212
219	# Build A3 + Ubuntu image	213	# Build A3 + Ubuntu image
220	-bash build.sh -t A3 -o ubuntu22.04	214	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-ubuntu22.04-py3.11
221		215
222	# Build with specified PyTorch version	216	# Build with specified PyTorch version
223	-bash build.sh -t A3 --torch-version 2.7.1 --torch-npu-version 2.7.1	217	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 --torch-version 2.7.1 --torch-npu-version 2.7.1.post4
224		218
225	# Build with offline .whl files	219	# Build with offline .whl files
226	-bash build.sh -t A3 \	220	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 \
227	--torch-whl /path/to/torch-2.7.1+cpu-cp311-cp311-linux_x86_64.whl \	221	--torch-whl /path/to/torch-2.7.1+cpu-cp311-cp311-linux_x86_64.whl \
228	--torch-npu-whl /path/to/torch_npu-2.7.1-cp311-cp311-linux_x86_64.whl	222	--torch-npu-whl /path/to/torch_npu-2.7.1-cp311-cp311-linux_x86_64.whl
229		223
230	-# Build with specified base image version
231	-bash build.sh -t A3 --base-image-version 9.0.0
232	-
233	# Build with specified version	224	# Build with specified version
234	-bash build.sh -t A3 -v 26.1.0	225	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -v 26.1.0
235	```	226	```
236		227
237	### Automatic Download Function Description	228	### Automatic Download Function Description
	@@ -246,12 +237,13 @@ The build script supports automatic download of the following resources. Please
246		237
247	### Automatic Base Image Recognition	238	### Automatic Base Image Recognition
248		239
249	-The build script automatically recognizes key information from the base image name:	240	+The build script automatically recognizes key information from the base image tag (the `--base-image` parameter is required):
250		241
251	-1. NPU type recognition: Recognizes the NPU type (e.g., `a3`) from the image tag	242	+1. CANN version recognition: Extracts the CANN version number (e.g., `9.0.0`) from the image tag
252	-2. Operating system recognition: Recognizes `openeuler24.03` or `ubuntu22.04` from the image tag	243	+2. NPU type recognition: Recognizes the NPU type (e.g., `a3`) from the image tag
253	-3. CANN version recognition: Extracts the CANN version number (e.g., `9.0.0`) from the image tag	244	+3. Operating system recognition: Recognizes `openeuler24.03` or `ubuntu22.04` from the image tag
254	-4. Automatic image tag generation: Automatically generates image tags conforming to naming rules based on recognized information	245	+4. Python version recognition: Extracts the Python version (e.g., `3.11`) from the `py<x.y>` field in the image tag
		246	+5. Automatic image tag generation: Automatically generates image tags conforming to naming rules based on recognized information
255		247
256	### Best Practice Example	248	### Best Practice Example
257		249
	@@ -262,7 +254,7 @@ The following example shows how to build a MindSpeed MM image using a custom bas
262	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"	254	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
263		255
264	# Custom configuration	256	# Custom configuration
265	-BASE_IMAGE="swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11"	257	+BASE_IMAGE="swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11"
266	TORCH_VERSION="2.7.1"	258	TORCH_VERSION="2.7.1"
267	TORCH_NPU_VERSION="2.7.1.post6"	259	TORCH_NPU_VERSION="2.7.1.post6"
268	MINDSPEED_MM_VERSION="26.0.0"	260	MINDSPEED_MM_VERSION="26.0.0"
	@@ -278,7 +270,7 @@ bash "${SCRIPT_DIR}/build.sh" \
278		270
279	Key Feature Description:	271	Key Feature Description:
280		272
281	-1. Automatic recognition: The script automatically recognizes the NPU type, operating system (openeuler24.03), and CANN version (9.0.0) from `BASE_IMAGE`. If `BASE_IMAGE` doesn't exist in the system, it will be automatically pulled.	273	+1. Automatic recognition: The script automatically recognizes the CANN version (9.0.0), NPU type (a3), operating system (openeuler24.03), and Python version (3.11) from `BASE_IMAGE`. If `BASE_IMAGE` doesn't exist in the system, it will be automatically pulled.
282	2. Automatic tag generation: Automatically generates image tags conforming to the naming rules based on recognition results.	274	2. Automatic tag generation: Automatically generates image tags conforming to the naming rules based on recognition results.
283	3. Automatic download: If the Miniconda installer or decord dependencies are not available locally, the script will automatically download them	275	3. Automatic download: If the Miniconda installer or decord dependencies are not available locally, the script will automatically download them
284	4. Failure cleanup: The `--cleanup-on-fail` parameter ensures cleanup of dangling resources if the build fails	276	4. Failure cleanup: The `--cleanup-on-fail` parameter ensures cleanup of dangling resources if the build fails
	@@ -375,7 +367,7 @@ Build an image containing the new model environment using the following command:
375		367
376	```bash	368	```bash
377	# Use the -v parameter to specify the version number, and the build script will automatically find the corresponding installation script	369	# Use the -v parameter to specify the version number, and the build script will automatically find the corresponding installation script
378	-bash build.sh -t A3 -v 26.1.0	370	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -v 26.1.0
379	```	371	```
380		372
381	#### 5. Update Documentation	373	#### 5. Update Documentation

Mdocker/OVERVIEW.zh.md+19-27

	@@ -66,7 +66,7 @@ MindSpeed MM 镜像基于 Ubuntu 22.04 和 openEuler 24.03 两种操作系统，
66	\| Python版本 \| `py3.11` \| Python 3.11 \|	66	\| Python版本 \| `py3.11` \| Python 3.11 \|
67	\| 架构类型 \| `aarch64` \| ARM64 架构 \|	67	\| 架构类型 \| `aarch64` \| ARM64 架构 \|
68		68
69	-> CANN 版本与 TorchNPU 版本来源于构建时的基础镜像和 `--torch-npu-version` 参数。	69	+> CANN 版本、NPU 类型、操作系统和 Python 版本均来源于构建时基础镜像的 tag，TorchNPU 版本来源于 `--torch-npu-version` 参数。
70		70
71	Dockerfile 命名遵循模板：`Dockerfile[.{芯片信息}.{操作系统}.{其他字段}]`	71	Dockerfile 命名遵循模板：`Dockerfile[.{芯片信息}.{操作系统}.{其他字段}]`
72		72
	@@ -181,12 +181,11 @@ docker run -it --rm \
181		181
182	### 构建脚本参数说明	182	### 构建脚本参数说明
183		183
184	-构建脚本 `build.sh` 支持多种参数配置，以下默认值仅为示例，请根据实际需求调整。其中 `--base-image-version` 的默认值 `9.0.0` 仅为示例，强烈建议通过 `--base-image` 或 `--base-image-version` 显式指定与目标环境匹配的基础镜像。	184	+构建脚本 `build.sh` 支持多种参数配置。其中 CANN 版本、NPU 类型、操作系统和 Python 版本只能通过解析 `--base-image` 的 tag 获取，不支持手动指定。
185		185
186	-\| 参数 \| 说明 \| 默认值（示例） \|	186	+\| 参数 \| 说明 \| 默认值 \|
187	\| ------ \| ------ \| ------------ \|	187	\| ------ \| ------ \| ------------ \|
188	-\| `-t, --npu-type` \| NPU 类型：A3 或 910B（未指定时可从 `--base-image` 自动识别） \| 无（必需） \|	188	+\| `--base-image` \| 必选。完整基础镜像名称，CANN 版本、NPU 类型、操作系统和 Python 版本均从镜像 tag 自动识别 \| 无（必需） \|
189	-\| `-o, --os` \| 操作系统：openeuler24.03 或 ubuntu22.04（未指定时可从 `--base-image` 自动识别） \| openeuler24.03 \|
190	\| `-v, --version` \| MindSpeed MM 版本标识，同时作为 Git 分支名称和脚本目录选择依据 \| 26.1.0 \|	189	\| `-v, --version` \| MindSpeed MM 版本标识，同时作为 Git 分支名称和脚本目录选择依据 \| 26.1.0 \|
191	\| `-m, --miniconda` \| Miniconda 安装器路径（未指定时自动下载） \| 无（自动下载） \|	190	\| `-m, --miniconda` \| Miniconda 安装器路径（未指定时自动下载） \| 无（自动下载） \|
192	\| `-d, --decord-deps` \| decord 依赖目录路径（ARM 架构下未指定时自动下载） \| 无（自动下载） \|	191	\| `-d, --decord-deps` \| decord 依赖目录路径（ARM 架构下未指定时自动下载） \| 无（自动下载） \|
	@@ -200,8 +199,6 @@ docker run -it --rm \
200	\| `--torch-npu-whl` \| torch-npu.whl 文件路径（离线安装） \| 无 \|	199	\| `--torch-npu-whl` \| torch-npu.whl 文件路径（离线安装） \| 无 \|
201	\| `--torchvision-whl` \| torchvision.whl 文件路径（可选，离线安装） \| 无 \|	200	\| `--torchvision-whl` \| torchvision.whl 文件路径（可选，离线安装） \| 无 \|
202	\| `--torchaudio-whl` \| torchaudio.whl 文件路径（可选，离线安装） \| 无 \|	201	\| `--torchaudio-whl` \| torchaudio.whl 文件路径（可选，离线安装） \| 无 \|
203	-\| `--base-image-version` \| 基础镜像 CANN 版本（示例值，建议显式指定） \| 9.0.0 \|
204	-\| `--base-image` \| 完整基础镜像名称（优先级高于 `--base-image-version`，推荐使用） \| 无 \|
205	\| `--build-ci` \| 构建 CI 镜像，包含多版本 conda 环境（跳过 verl + MindSpeed-MM 克隆） \| 无 \|	202	\| `--build-ci` \| 构建 CI 镜像，包含多版本 conda 环境（跳过 verl + MindSpeed-MM 克隆） \| 无 \|
206	\| `--cleanup-on-fail` \| 构建失败时清理悬空镜像/容器 \| 无 \|	203	\| `--cleanup-on-fail` \| 构建失败时清理悬空镜像/容器 \| 无 \|
207		204
	@@ -210,28 +207,22 @@ docker run -it --rm \
210	```bash	207	```bash
211	cd docker	208	cd docker
212		209
213	-# 构建 A3 + openEuler 镜像（默认）	210	+# 构建 A3 + openEuler 镜像
214	-bash build.sh -t A3	211	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11
215	-
216	-# 构建 A2 + openEuler 镜像
217	-bash build.sh -t 910B
218		212
219	# 构建 A3 + Ubuntu 镜像	213	# 构建 A3 + Ubuntu 镜像
220	-bash build.sh -t A3 -o ubuntu22.04	214	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-ubuntu22.04-py3.11
221		215
222	# 指定 PyTorch 版本构建	216	# 指定 PyTorch 版本构建
223	-bash build.sh -t A3 --torch-version 2.7.1 --torch-npu-version 2.7.1	217	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 --torch-version 2.7.1 --torch-npu-version 2.7.1.post4
224		218
225	# 使用离线 .whl 文件构建	219	# 使用离线 .whl 文件构建
226	-bash build.sh -t A3 \	220	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 \
227	--torch-whl /path/to/torch-2.7.1+cpu-cp311-cp311-linux_x86_64.whl \	221	--torch-whl /path/to/torch-2.7.1+cpu-cp311-cp311-linux_x86_64.whl \
228	--torch-npu-whl /path/to/torch_npu-2.7.1-cp311-cp311-linux_x86_64.whl	222	--torch-npu-whl /path/to/torch_npu-2.7.1-cp311-cp311-linux_x86_64.whl
229		223
230	-# 指定基础镜像版本构建
231	-bash build.sh -t A3 --base-image-version 9.0.0
232	-
233	# 指定版本构建	224	# 指定版本构建
234	-bash build.sh -t A3 -v 26.1.0	225	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -v 26.1.0
235	```	226	```
236		227
237	### 自动下载功能说明	228	### 自动下载功能说明
	@@ -246,12 +237,13 @@ bash build.sh -t A3 -v 26.1.0
246		237
247	### 自动识别基础镜像	238	### 自动识别基础镜像
248		239
249	-构建脚本会自动识别基础镜像名称中的关键信息：	240	+构建脚本会自动从基础镜像 tag 中识别关键信息（`--base-image` 参数为必选）：
250		241
251	-1. NPU 类型识别：从镜像 tag 中识别NPU类型(如`a3`)	242	+1. CANN 版本识别：从镜像 tag 中提取 CANN 版本号（如 `9.0.0`）
252	-2. 操作系统识别：从镜像 tag 中识别 `openeuler24.03` 或 `ubuntu22.04`	243	+2. NPU 类型识别：从镜像 tag 中识别 NPU 类型（如 `a3`）
253	-3. CANN 版本识别：从镜像 tag 中提取 CANN 版本号（如 `9.0.0`）	244	+3. 操作系统识别：从镜像 tag 中识别 `openeuler24.03` 或 `ubuntu22.04`
254	-4. 自动生成镜像 tag：基于识别到的信息自动生成符合命名规则的镜像 tag	245	+4. Python 版本识别：从镜像 tag 中的 `py<x.y>` 字段提取 Python 版本（如 `3.11`）
		246	+5. 自动生成镜像 tag：基于识别到的信息自动生成符合命名规则的镜像 tag
255		247
256	### 最佳实践示例	248	### 最佳实践示例
257		249
	@@ -262,7 +254,7 @@ bash build.sh -t A3 -v 26.1.0
262	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"	254	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
263		255
264	# 自定义配置	256	# 自定义配置
265	-BASE_IMAGE="swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11"	257	+BASE_IMAGE="swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11"
266	TORCH_VERSION="2.7.1"	258	TORCH_VERSION="2.7.1"
267	TORCH_NPU_VERSION="2.7.1.post6"	259	TORCH_NPU_VERSION="2.7.1.post6"
268	MINDSPEED_MM_VERSION="26.0.0"	260	MINDSPEED_MM_VERSION="26.0.0"
	@@ -278,7 +270,7 @@ bash "${SCRIPT_DIR}/build.sh" \
278		270
279	关键特性说明：	271	关键特性说明：
280		272
281	-1. 自动识别：脚本会自动从 `BASE_IMAGE` 中识别 NPU 类型、操作系统（openeuler24.03）和 CANN 版本（9.0.0）。如果`BASE_IMAGE`在系统中不存在，会自动拉取。	273	+1. 自动识别：脚本会自动从 `BASE_IMAGE` 中识别 CANN 版本（9.0.0）、NPU 类型（a3）、操作系统（openeuler24.03）和 Python 版本（3.11）。如果`BASE_IMAGE`在系统中不存在，会自动拉取。
282	2. 自动生成 tag：基于识别结果自动生成符合命名规则的镜像 tag。	274	2. 自动生成 tag：基于识别结果自动生成符合命名规则的镜像 tag。
283	3. 自动下载：如果本地没有 Miniconda 安装器或 decord 依赖，脚本会自动下载	275	3. 自动下载：如果本地没有 Miniconda 安装器或 decord 依赖，脚本会自动下载
284	4. 失败清理：`--cleanup-on-fail` 参数确保构建失败时清理悬空资源	276	4. 失败清理：`--cleanup-on-fail` 参数确保构建失败时清理悬空资源
	@@ -375,7 +367,7 @@ echo "=========================================="
375		367
376	```bash	368	```bash
377	# 使用 -v 参数指定版本号，构建脚本会自动找到对应的安装脚本	369	# 使用 -v 参数指定版本号，构建脚本会自动找到对应的安装脚本
378	-bash build.sh -t A3 -v 26.1.0	370	+bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -v 26.1.0
379	```	371	```
380		372
381	#### 5. 更新文档	373	#### 5. 更新文档

Mdocker/build.sh+73-73

	@@ -26,7 +26,9 @@ cleanup_dangling() {
26	# Dockerfile naming: Dockerfile (unified, supports all NPU types and OS)	26	# Dockerfile naming: Dockerfile (unified, supports all NPU types and OS)
27	# Image tag naming: {version}-cann{cann_version}-torch_npu{torch_npu_version}-{chip}-{os}-py{python_version}-{architecture}	27	# Image tag naming: {version}-cann{cann_version}-torch_npu{torch_npu_version}-{chip}-{os}-py{python_version}-{architecture}
28	# Usage:	28	# Usage:
29	-# bash build.sh -t A3 [-m /path/to/miniconda.sh] [-o ubuntu22.04]	29	+# bash build.sh --base-image <base-image> [-m /path/to/miniconda.sh]
		30	+# CANN version, NPU type, OS and Python version are auto-detected from the
		31	+# base image tag and cannot be specified manually.
30	# ============================================	32	# ============================================
31		33
32	set -e	34	set -e
	@@ -42,7 +44,8 @@ Usage: $0 [OPTIONS]
42	Build MindSpeed MM Docker Image	44	Build MindSpeed MM Docker Image
43		45
44	Required:	46	Required:
45	- -t, --npu-type TYPE NPU type: A3 or 910B (auto-detected from --base-image if not specified)	47	+ --base-image IMAGE Full base image name (e.g. swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11)
		48	+ CANN version, NPU type, OS and Python version are auto-detected from the image tag.
46		49
47	Optional:	50	Optional:
48	-m, --miniconda PATH Miniconda installer path (auto-download if not specified)	51	-m, --miniconda PATH Miniconda installer path (auto-download if not specified)
	@@ -51,7 +54,6 @@ Optional:
51	-i, --image-name NAME Image name (default: mindspeed-mm:v{version}-cann{cann_ver}-torch_npu{torch_npu_ver}-{chip}-{os}-py{py_ver}-{arch})	54	-i, --image-name NAME Image name (default: mindspeed-mm:v{version}-cann{cann_ver}-torch_npu{torch_npu_ver}-{chip}-{os}-py{py_ver}-{arch})
52	--tag TAG Custom image tag (overrides the default tag; keeps repo name 'mindspeed-mm')	55	--tag TAG Custom image tag (overrides the default tag; keeps repo name 'mindspeed-mm')
53	-n, --no-cache Build without cache	56	-n, --no-cache Build without cache
54	- -o, --os OS OS: openeuler24.03 or ubuntu22.04 (auto-detected from --base-image if not specified)
55	-v, --version VERSION MindSpeed MM version (default: 26.1.0, determines model install scripts)	57	-v, --version VERSION MindSpeed MM version (default: 26.1.0, determines model install scripts)
56	--torch-version VER PyTorch version (default: 2.7.1, for online install)	58	--torch-version VER PyTorch version (default: 2.7.1, for online install)
57	--torch-npu-version VER torch-npu version (default: 2.7.1, for online install)	59	--torch-npu-version VER torch-npu version (default: 2.7.1, for online install)
	@@ -59,14 +61,11 @@ Optional:
59	--torch-npu-whl PATH torch-npu .whl file path (offline install)	61	--torch-npu-whl PATH torch-npu .whl file path (offline install)
60	--torchvision-whl PATH torchvision .whl file path (optional, offline install)	62	--torchvision-whl PATH torchvision .whl file path (optional, offline install)
61	--torchaudio-whl PATH torchaudio .whl file path (optional, offline install)	63	--torchaudio-whl PATH torchaudio .whl file path (optional, offline install)
62	- --base-image-version VER Base image CANN version (default: 9.0.0, example only; recommended to specify via --base-image or --base-image-version)
63	- --base-image IMAGE Full base image name (higher priority than --base-image-version; recommended)
64	--build-ci Build CI image with multi-version conda environments (skip verl + MindSpeed-MM clone)	64	--build-ci Build CI image with multi-version conda environments (skip verl + MindSpeed-MM clone)
65	--cleanup-on-fail Clean up dangling images/containers if build fails	65	--cleanup-on-fail Clean up dangling images/containers if build fails
66	-h, --help Show help	66	-h, --help Show help
67		67
68	Dockerfile naming convention: Dockerfile (unified, supports all NPU types and OS)	68	Dockerfile naming convention: Dockerfile (unified, supports all NPU types and OS)
69	- NPU type and OS are passed as build arguments
70		69
71	Image tag naming convention: {version}-cann{cann_version}-torch_npu{torch_npu_version}-{chip}-{os}-py{python_version}-{architecture}[-ci]	70	Image tag naming convention: {version}-cann{cann_version}-torch_npu{torch_npu_version}-{chip}-{os}-py{python_version}-{architecture}[-ci]
72	e.g. v26.0.0-cann9.0.0-torch_npu2.7.1-a3-openeuler24.03-py3.11-x86_64	71	e.g. v26.0.0-cann9.0.0-torch_npu2.7.1-a3-openeuler24.03-py3.11-x86_64
	@@ -74,51 +73,39 @@ Image tag naming convention: {version}-cann{cann_version}-torch_npu{torch_npu_ve
74	v26.1.0-cann9.0.0-torch_npu2.7.1-910b-openeuler24.03-py3.11-aarch64-ci (CI build via --build-ci)	73	v26.1.0-cann9.0.0-torch_npu2.7.1-910b-openeuler24.03-py3.11-aarch64-ci (CI build via --build-ci)
75		74
76	Examples:	75	Examples:
77	- bash $0 -t A3	76	+ bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11
78	- bash $0 -t A3 -o ubuntu22.04	77	+ bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11 --torch-version 2.7.1 --torch-npu-version 2.7.1
79	- bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-beta.2-910b-openeuler24.03-py3.11	78	+ bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -i myproject/mindspeed-mm:v1.0
80	- bash $0 -t 910B --torch-version 2.7.1 --torch-npu-version 2.7.1	79	+ bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 --torch-whl /path/to/torch.whl --torch-npu-whl /path/to/torch_npu.whl
81	- bash $0 -t A3 --base-image-version 9.0.0	80	+ bash $0 --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11 --build-ci # Build CI image (tag auto-appends '-ci')
82	- bash $0 -t A3 -i myproject/mindspeed-mm:v1.0
83	- bash $0 -t A3 --torch-whl /path/to/torch.whl --torch-npu-whl /path/to/torch_npu.whl
84	- bash $0 -t 910B --build-ci # Build CI image (tag auto-appends '-ci')
85	EOF	81	EOF
86	}	82	}
87		83
88	-NPU_TYPE=""
89	MINICONDA_PATH=""	84	MINICONDA_PATH=""
90	DECORD_DEPS_PATH=""	85	DECORD_DEPS_PATH=""
91	DECORD_SCRIPT_PATH=""	86	DECORD_SCRIPT_PATH=""
92	IMAGE_NAME=""	87	IMAGE_NAME=""
93	IMAGE_TAG=""	88	IMAGE_TAG=""
94	NO_CACHE=""	89	NO_CACHE=""
95	-OS="openeuler24.03"
96	TORCH_VERSION="2.7.1"	90	TORCH_VERSION="2.7.1"
97	TORCH_NPU_VERSION="2.7.1"	91	TORCH_NPU_VERSION="2.7.1"
98	TORCH_WHL_PATH=""	92	TORCH_WHL_PATH=""
99	TORCH_NPU_WHL_PATH=""	93	TORCH_NPU_WHL_PATH=""
100	TORCHVISION_WHL_PATH=""	94	TORCHVISION_WHL_PATH=""
101	TORCHAUDIO_WHL_PATH=""	95	TORCHAUDIO_WHL_PATH=""
102	-# Default value is an example only. It is recommended to explicitly specify the	96	+BASE_IMAGE=""
103	-# base image via --base-image or --base-image-version to match your target environment.
104	-BASE_IMAGE_VERSION="9.0.0"
105	MINDSPEED_MM_VERSION="26.1.0"	97	MINDSPEED_MM_VERSION="26.1.0"
106	-PYTHON_VERSION="3.11"
107	-NPU_TYPE_EXPLICIT=false
108	-OS_EXPLICIT=false
109	CLEANUP_ON_FAIL=false	98	CLEANUP_ON_FAIL=false
110	BUILD_CI=false	99	BUILD_CI=false
111		100
112	while [[ $# -gt 0 ]]; do	101	while [[ $# -gt 0 ]]; do
113	case $1 in	102	case $1 in
114	- -t\|--npu-type) NPU_TYPE="$2"; NPU_TYPE_EXPLICIT=true; shift 2 ;;
115	-m\|--miniconda) MINICONDA_PATH="$2"; shift 2 ;;	103	-m\|--miniconda) MINICONDA_PATH="$2"; shift 2 ;;
116	-d\|--decord-deps) DECORD_DEPS_PATH="$2"; shift 2 ;;	104	-d\|--decord-deps) DECORD_DEPS_PATH="$2"; shift 2 ;;
117	-s\|--decord-script) DECORD_SCRIPT_PATH="$2"; shift 2 ;;	105	-s\|--decord-script) DECORD_SCRIPT_PATH="$2"; shift 2 ;;
118	-i\|--image-name) IMAGE_NAME="$2"; shift 2 ;;	106	-i\|--image-name) IMAGE_NAME="$2"; shift 2 ;;
119	--tag) IMAGE_TAG="$2"; shift 2 ;;	107	--tag) IMAGE_TAG="$2"; shift 2 ;;
120	-n\|--no-cache) NO_CACHE="--no-cache"; shift ;;	108	-n\|--no-cache) NO_CACHE="--no-cache"; shift ;;
121	- -o\|--os) OS="$2"; OS_EXPLICIT=true; shift 2 ;;
122	-v\|--version) MINDSPEED_MM_VERSION="$2"; shift 2 ;;	109	-v\|--version) MINDSPEED_MM_VERSION="$2"; shift 2 ;;
123	--torch-version) TORCH_VERSION="$2"; shift 2 ;;	110	--torch-version) TORCH_VERSION="$2"; shift 2 ;;
124	--torch-npu-version) TORCH_NPU_VERSION="$2"; shift 2 ;;	111	--torch-npu-version) TORCH_NPU_VERSION="$2"; shift 2 ;;
	@@ -126,7 +113,6 @@ while [[ $# -gt 0 ]]; do
126	--torch-npu-whl) TORCH_NPU_WHL_PATH="$2"; shift 2 ;;	113	--torch-npu-whl) TORCH_NPU_WHL_PATH="$2"; shift 2 ;;
127	--torchvision-whl) TORCHVISION_WHL_PATH="$2"; shift 2 ;;	114	--torchvision-whl) TORCHVISION_WHL_PATH="$2"; shift 2 ;;
128	--torchaudio-whl) TORCHAUDIO_WHL_PATH="$2"; shift 2 ;;	115	--torchaudio-whl) TORCHAUDIO_WHL_PATH="$2"; shift 2 ;;
129	- --base-image-version) BASE_IMAGE_VERSION="$2"; shift 2 ;;
130	--base-image) BASE_IMAGE="$2"; shift 2 ;;	116	--base-image) BASE_IMAGE="$2"; shift 2 ;;
131	--build-ci) BUILD_CI=true; shift ;;	117	--build-ci) BUILD_CI=true; shift ;;
132	--cleanup-on-fail) CLEANUP_ON_FAIL=true; shift ;;	118	--cleanup-on-fail) CLEANUP_ON_FAIL=true; shift ;;
	@@ -172,6 +158,12 @@ parse_base_image_tag() {
172	detected_cann="${tag_lower%%-*}"	158	detected_cann="${tag_lower%%-*}"
173	fi	159	fi
174		160
		161	+ # Python version: match 'py<x.y>' field in the tag (e.g. py3.11 -> 3.11)
		162	+ local detected_python=""
		163	+ if [[ "$tag_lower" =~ py([0-9]+\.[0-9]+) ]]; then
		164	+ detected_python="${BASH_REMATCH[1]}"
		165	+ fi
		166	+
175	if [ -n "$detected_npu" ]; then	167	if [ -n "$detected_npu" ]; then
176	DETECTED_NPU_TYPE="$detected_npu"	168	DETECTED_NPU_TYPE="$detected_npu"
177	echo ">>> Auto-detected NPU type from base image: $detected_npu"	169	echo ">>> Auto-detected NPU type from base image: $detected_npu"
	@@ -187,34 +179,57 @@ parse_base_image_tag() {
187	echo ">>> Auto-detected OS from base image: $detected_os"	179	echo ">>> Auto-detected OS from base image: $detected_os"
188	fi	180	fi
189		181
		182	+ if [ -n "$detected_python" ]; then
		183	+ DETECTED_PYTHON_VERSION="$detected_python"
		184	+ echo ">>> Auto-detected Python version from base image: $detected_python"
		185	+ fi
		186	+
190	return 0	187	return 0
191	}	188	}
192		189
193	DETECTED_NPU_TYPE=""	190	DETECTED_NPU_TYPE=""
194	DETECTED_OS=""	191	DETECTED_OS=""
195	DETECTED_CANN_VERSION=""	192	DETECTED_CANN_VERSION=""
		193	+DETECTED_PYTHON_VERSION=""
196		194
197	-if [ -n "$BASE_IMAGE" ]; then	195	+if [ -z "$BASE_IMAGE" ]; then
198	- echo ">>> Auto-detecting NPU type, OS, and CANN version from base image..."	196	+ echo "Error: --base-image is required. CANN version, NPU type, OS and Python version are auto-detected from the base image tag."
199	- parse_base_image_tag "$BASE_IMAGE"
200	-
201	- if [ "$NPU_TYPE_EXPLICIT" = false ] && [ -n "$DETECTED_NPU_TYPE" ]; then
202	- NPU_TYPE="$DETECTED_NPU_TYPE"
203	- echo ">>> Using auto-detected NPU type: $NPU_TYPE"
204	- fi
205	-
206	- if [ "$OS_EXPLICIT" = false ] && [ -n "$DETECTED_OS" ]; then
207	- OS="$DETECTED_OS"
208	- echo ">>> Using auto-detected OS: $OS"
209	- fi
210	-fi
211	-
212	-if [ -z "$NPU_TYPE" ]; then
213	- echo "Error: NPU type is required (-t or --npu-type) or provide --base-image for auto-detection"
214	show_help	197	show_help
215	exit 1	198	exit 1
216	fi	199	fi
217		200
		201	+echo ">>> Auto-detecting NPU type, OS, CANN version and Python version from base image..."
		202	+parse_base_image_tag "$BASE_IMAGE"
		203	+
		204	+if [ -z "$DETECTED_NPU_TYPE" ]; then
		205	+ echo "Error: Failed to detect NPU type from base image tag (expected '910b' or 'a3')."
		206	+ exit 1
		207	+fi
		208	+NPU_TYPE="$DETECTED_NPU_TYPE"
		209	+
		210	+if [ -z "$DETECTED_OS" ]; then
		211	+ echo "Error: Failed to detect OS from base image tag (expected 'openeuler24.03' or 'ubuntu22.04')."
		212	+ exit 1
		213	+fi
		214	+OS="$DETECTED_OS"
		215	+
		216	+if [ -z "$DETECTED_CANN_VERSION" ]; then
		217	+ echo "Error: Failed to detect CANN version from base image tag."
		218	+ exit 1
		219	+fi
		220	+CANN_VERSION="$DETECTED_CANN_VERSION"
		221	+
		222	+if [ -z "$DETECTED_PYTHON_VERSION" ]; then
		223	+ echo "Error: Failed to detect Python version from base image tag (expected 'py<x.y>' field, e.g. py3.11)."
		224	+ exit 1
		225	+fi
		226	+PYTHON_VERSION="$DETECTED_PYTHON_VERSION"
		227	+
		228	+echo ">>> Using detected NPU type: $NPU_TYPE"
		229	+echo ">>> Using detected OS: $OS"
		230	+echo ">>> Using detected CANN version: $CANN_VERSION"
		231	+echo ">>> Using detected Python: $PYTHON_VERSION"
		232	+
218	NPU_TYPE=$(echo "$NPU_TYPE" \| tr '[:lower:]' '[:upper:]')	233	NPU_TYPE=$(echo "$NPU_TYPE" \| tr '[:lower:]' '[:upper:]')
219	NPU_TYPE_LOWER=$(echo "$NPU_TYPE" \| tr '[:upper:]' '[:lower:]')	234	NPU_TYPE_LOWER=$(echo "$NPU_TYPE" \| tr '[:upper:]' '[:lower:]')
220	OS=$(echo "$OS" \| tr '[:upper:]' '[:lower:]')	235	OS=$(echo "$OS" \| tr '[:upper:]' '[:lower:]')
	@@ -319,15 +334,6 @@ if [[ "$MINICONDA_NAME" == "aarch64" ]]; then
319	ARCH_NAME="aarch64"	334	ARCH_NAME="aarch64"
320	fi	335	fi
321		336
322	-# Determine CANN version for the image tag:
323	-# - From the base image tag (auto-detected) when --base-image is provided
324	-# - From --base-image-version otherwise
325	-if [ -n "$BASE_IMAGE" ] && [ -n "$DETECTED_CANN_VERSION" ]; then
326	- CANN_VERSION="$DETECTED_CANN_VERSION"
327	-else
328	- CANN_VERSION="$BASE_IMAGE_VERSION"
329	-fi
330	-
331	if [ -n "$IMAGE_NAME" ] && [ -n "$IMAGE_TAG" ]; then	337	if [ -n "$IMAGE_NAME" ] && [ -n "$IMAGE_TAG" ]; then
332	echo "Warning: both --image-name and --tag provided; --image-name takes precedence (--tag ignored)"	338	echo "Warning: both --image-name and --tag provided; --image-name takes precedence (--tag ignored)"
333	fi	339	fi
	@@ -404,14 +410,15 @@ fi
404	echo "=========================================="	410	echo "=========================================="
405	echo "Build Configuration"	411	echo "Build Configuration"
406	echo "=========================================="	412	echo "=========================================="
		413	+echo "Base Image: ${BASE_IMAGE}"
407	echo "NPU Type: ${NPU_TYPE}"	414	echo "NPU Type: ${NPU_TYPE}"
408	echo "OS: ${OS}"	415	echo "OS: ${OS}"
		416	+echo "CANN Version: ${CANN_VERSION}"
		417	+echo "Python Version: ${PYTHON_VERSION}"
409	echo "CPU Architecture: ${ARCH_NAME}"	418	echo "CPU Architecture: ${ARCH_NAME}"
410	echo "Dockerfile: ${DOCKERFILE}"	419	echo "Dockerfile: ${DOCKERFILE}"
411	echo "Image Name: ${IMAGE_NAME}"	420	echo "Image Name: ${IMAGE_NAME}"
412	echo "Image Tag: ${IMAGE_TAG:-(auto)}"	421	echo "Image Tag: ${IMAGE_TAG:-(auto)}"
413	-echo "Base Image Version: ${BASE_IMAGE_VERSION}"
414	-echo "CANN Version: ${CANN_VERSION}"
415	echo "PyTorch Version: ${TORCH_VERSION}"	422	echo "PyTorch Version: ${TORCH_VERSION}"
416	echo "torch-npu Version: ${TORCH_NPU_VERSION}"	423	echo "torch-npu Version: ${TORCH_NPU_VERSION}"
417	echo "MindSpeed MM Ver: ${MINDSPEED_MM_VERSION}"	424	echo "MindSpeed MM Ver: ${MINDSPEED_MM_VERSION}"
	@@ -434,21 +441,19 @@ echo "Build CI: ${BUILD_CI}"
434	echo "No Cache: ${NO_CACHE:-No}"	441	echo "No Cache: ${NO_CACHE:-No}"
435	echo "=========================================="	442	echo "=========================================="
436		443
437	-if [ -n "$BASE_IMAGE" ]; then	444	+echo ""
438	- echo ""	445	+echo ">>> Checking if base image exists..."
439	- echo ">>> Checking if base image exists..."	446	+if ! docker image inspect "$BASE_IMAGE" > /dev/null 2>&1; then
440	- if ! docker image inspect "$BASE_IMAGE" > /dev/null 2>&1; then	447	+ echo ">>> Base image not found, pulling: ${BASE_IMAGE}"
441	- echo ">>> Base image not found, pulling: ${BASE_IMAGE}"	448	+ docker pull "$BASE_IMAGE"
442	- docker pull "$BASE_IMAGE"	449	+ if [ $? -ne 0 ]; then
443	- if [ $? -ne 0 ]; then	450	+ echo "Error: Failed to pull base image"
444	- echo "Error: Failed to pull base image"	451	+ exit 1
445	- exit 1
446	- fi
447	- else
448	- echo ">>> Base image already exists: ${BASE_IMAGE}"
449	fi	452	fi
450	- echo ""	453	+else
		454	+ echo ">>> Base image already exists: ${BASE_IMAGE}"
451	fi	455	fi
		456	+echo ""
452		457
453	cd "$SCRIPT_DIR"	458	cd "$SCRIPT_DIR"
454		459
	@@ -506,9 +511,7 @@ fi
506		511
507	BUILD_ARGS="--build-arg MINICONDA_SH=${MINICONDA_NAME}"	512	BUILD_ARGS="--build-arg MINICONDA_SH=${MINICONDA_NAME}"
508	BUILD_ARGS="$BUILD_ARGS --build-arg DECORD_SCRIPT=${DECORD_SCRIPT_NAME}"	513	BUILD_ARGS="$BUILD_ARGS --build-arg DECORD_SCRIPT=${DECORD_SCRIPT_NAME}"
509	-BUILD_ARGS="$BUILD_ARGS --build-arg OS=${OS}"
510	BUILD_ARGS="$BUILD_ARGS --build-arg OS_FAMILY=${OS_FAMILY}"	514	BUILD_ARGS="$BUILD_ARGS --build-arg OS_FAMILY=${OS_FAMILY}"
511	-BUILD_ARGS="$BUILD_ARGS --build-arg NPU_TYPE=${NPU_TYPE_LOWER}"
512	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_VERSION=${TORCH_VERSION}"	515	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_VERSION=${TORCH_VERSION}"
513	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_NPU_VERSION=${TORCH_NPU_VERSION}"	516	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_NPU_VERSION=${TORCH_NPU_VERSION}"
514	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_WHL_DIR=torch_wheels"	517	BUILD_ARGS="$BUILD_ARGS --build-arg TORCH_WHL_DIR=torch_wheels"
	@@ -519,11 +522,8 @@ if [ "$BUILD_CI" = true ]; then
519	BUILD_ARGS="$BUILD_ARGS --build-arg FINAL_WORKDIR=/workspace"	522	BUILD_ARGS="$BUILD_ARGS --build-arg FINAL_WORKDIR=/workspace"
520	fi	523	fi
521		524
522	-if [ -n "$BASE_IMAGE" ]; then	525	+# Always pass the full BASE_IMAGE to the Dockerfile (required, auto-detected values come from its tag).
523	- BUILD_ARGS="$BUILD_ARGS --build-arg BASE_IMAGE=${BASE_IMAGE}"	526	+BUILD_ARGS="$BUILD_ARGS --build-arg BASE_IMAGE=${BASE_IMAGE}"
524	-else
525	- BUILD_ARGS="$BUILD_ARGS --build-arg BASE_IMAGE_VERSION=${BASE_IMAGE_VERSION}"
526	-fi
527		527
528	BUILD_ARGS="$BUILD_ARGS --build-arg DECORD_DEPS_DIR=${DECORD_DEPS_NAME}"	528	BUILD_ARGS="$BUILD_ARGS --build-arg DECORD_DEPS_DIR=${DECORD_DEPS_NAME}"
529		529

Mdocker/common/download_miniconda.sh+1-1

	@@ -89,4 +89,4 @@ echo ""
89	echo "Next steps:"	89	echo "Next steps:"
90	echo " 1. Verify the file: bash ${MINICONDA_FILE} --help"	90	echo " 1. Verify the file: bash ${MINICONDA_FILE} --help"
91	echo " 2. Use it to build Docker image:"	91	echo " 2. Use it to build Docker image:"
92	-echo " bash build.sh -t A3 -m ${OUTPUT_DIR}/${MINICONDA_FILE}"	92	+echo " bash build.sh --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-a3-openeuler24.03-py3.11 -m ${OUTPUT_DIR}/${MINICONDA_FILE}"