Cosmos3 昇腾 NPU 适配使用指南【免费下载链接】cann-recipes-embodied-ai本项目针对具身智能业务中的典型模型、加速算法提供基于CANN平台的优化样例项目地址: https://gitcode.com/cann/cann-recipes-embodied-aiCosmos3 整体介绍功能介绍Cosmos3是一个世界基础模型World Foundation Models框架面向物理 AI、机器人、自动驾驶、视频生成与多模态理解等场景。Cosmos3-Nano 支持文生视频T2V、图生视频I2V、视频生视频V2V等推理任务并可结合结构化输入完成世界模型生成与理解。本样例基于 Cosmos3 框架完成昇腾 NPU 适配提供依赖配置、设备适配脚本、FIA attention 后端、本地权重路径适配和基础推理验证命令帮助用户在昇腾 A3 环境上运行 Cosmos3-Nano 推理。代码仓拉取与适配文件覆盖本样例基于 Cosmos3 框架进行昇腾 NPU 适配。使用时先拉取 CANN 适配仓和 原始代码仓再将world_model/cosmos3下的适配文件覆盖到 Cosmos3 仓根目录。# 进入需要放置代码仓的本地目录建议让 cann-recipes-embodied-ai 与 cosmos-framework 保持同级 git clone https://gitcode.com/cann/cann-recipes-embodied-ai.git git clone https://github.com/NVIDIA/cosmos-framework.git cd cosmos-framework git checkout a61b292 # 回到两个代码仓的共同上级目录 cd ../ # 将 CANN 仓中的 Cosmos3 适配文件覆盖到 Cosmos3 仓根目录 cp -rf cann-recipes-embodied-ai/world_model/cosmos3/* ./cosmos-framework完成覆盖后cosmos-framework根目录下应包含npu_adapt.sh、pyproject.toml等适配文件。Cosmos3 在昇腾 A3 上的运行环境配置与昇腾平台相关的环境配置安装 CANN 软件包。本样例依赖 CANN 开发套件包cann-toolkit与 CANN 二进制算子包cann-kernels支持的 CANN 软件版本为CANN 9.0.0torch_npu2.10.0,python3.13。请从软件包下载地址下载Ascend-cann-toolkit_${version}_linux-${aarch}.run与Atlas-A3-cann-kernels_${version}_linux-${aarch}.run软件包并参考 CANN 安装文档 依次进行安装。${version}表示 CANN 包版本号如 9.0.0${aarch}表示 CPU 架构如 aarch64、x86_64# ${cann_install_path} 为 CANN 包的实际安装目录注意每次新建终端时首先 source 一下 set_env.sh。 # 方式1默认路径安装以 root 用户为例 source /usr/local/Ascend/ascend-toolkit/set_env.sh # 方式2指定路径进行安装 source ${cann_install_path}/ascend-toolkit/set_env.shuv 环境管理工具安装可选如果当前环境已经安装 uv可以跳过wget -qO- https://astral.sh/uv/install.sh | shPython 运行环境安装cd cosmos-framework uv sync --python 3.13模型权重下载本样例使用 Cosmos3-Nano 权重进行推理验证。请从 nvidia/Cosmos3-Nano 下载模型权重并将推理命令中的COSMOS_CHECKPOINT指向本地权重目录。此外视频生成还需要 Wan2.2 VAE 权重。请从 Wan-AI/Wan2.2-TI2V-5B 下载Wan2.2_VAE.pth只需要该 VAE 权重文件无需下载完整 Wan2.2-TI2V-5B 模型并将其放置到 Cosmos3-Nano 本地权重目录下。# 示例将 Cosmos3-Nano 权重放置在本地目录 /mnt/workspace/cosmos3/cosmos3-nano export COSMOS_CHECKPOINT/mnt/workspace/cosmos3/cosmos3-nano # Wan2.2 VAE 权重应放置为如下路径 ls ${COSMOS_CHECKPOINT}/Wan2.2_VAE.pth如果部署环境无法直接访问 Hugging Face可在可联网环境下载完整 Cosmos3-Nano 权重目录和Wan2.2_VAE.pth后拷贝到昇腾服务器再使用本地路径作为--checkpoint-path。执行 NPU 适配脚本cd cosmos-framework bash npu_adapt.sh推理验证示例完成适配后可在cosmos-framework根目录下执行以下命令进行基础场景验证。COSMOS_CHECKPOINT用于指定本地权重目录如不设置可直接将命令中的--checkpoint-path替换为实际权重路径。export COSMOS_CHECKPOINT/mnt/workspace/cosmos3/cosmos3-nano export COSMOS_RESOLUTION480 export COSMOS_SEED0 export COSMOS_NPUS1输入 JSON 配置说明推理命令中的-i inputs/omni/t2v.json用于指定单条样例输入。常用字段如下model_mode任务类型例如text2video、image2video、video2video。prompt文本提示词T2V 只需要配置该字段即可。vision_pathI2V/V2V 的输入图片或视频路径仅图生视频、视频生视频需要。vision_path可以写远程 URL也可以写本地文件路径。若服务器无法访问 GitHub/Hugging Face或遇到证书、代理、内网限制等网络问题请先手动下载输入图片/视频到本地然后在 JSON 中改成本地绝对路径例如{ model_mode: image2video, prompt: A robot arm moves smoothly in a lab., vision_path: /mnt/workspace/cosmos3/inputs/robot_153.jpg }T2V 示例 JSON 可简化为{ model_mode: text2video, name: t2v, prompt: A realistic video of molten metal being poured in a steel mill. }T2V 文生视频torchrun --nproc-per-node${COSMOS_NPUS} -m cosmos_framework.scripts.inference \ --parallelism-presetlatency \ -i inputs/omni/t2v.json \ -o outputs/t2v \ --checkpoint-path ${COSMOS_CHECKPOINT} \ --resolution${COSMOS_RESOLUTION} \ --seed${COSMOS_SEED} \ --no-guardrailsI2V 图生视频torchrun --nproc-per-node${COSMOS_NPUS} -m cosmos_framework.scripts.inference \ --parallelism-presetlatency \ -i inputs/omni/i2v.json \ -o outputs/i2v \ --checkpoint-path ${COSMOS_CHECKPOINT} \ --resolution${COSMOS_RESOLUTION} \ --seed${COSMOS_SEED} \ --no-guardrailsV2V 视频生视频torchrun --nproc-per-node${COSMOS_NPUS} -m cosmos_framework.scripts.inference \ --parallelism-presetlatency \ -i inputs/omni/v2v.json \ -o outputs/v2v \ --checkpoint-path ${COSMOS_CHECKPOINT} \ --resolution${COSMOS_RESOLUTION} \ --seed${COSMOS_SEED} \ --no-guardrails样例输出展示以下为在昇腾 NPU 上运行上述基础场景得到的示例输出可用于快速查看生成效果。场景示例输出T2V 文生视频I2V 图生视频V2V 视频生视频citation## Citation misc{nvidia2026cosmos3omnimodalworldmodels, title{Cosmos 3: Omnimodal World Models for Physical AI}, author{NVIDIA: Aditi, Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, Aarti Basant, Mukesh Beladiya, Mohammad Qazim Bhat, Zaid Pervaiz Bhat, Dan Blick, Vanni Brighella, Han Cai, Tiffany Cai, Eric Cameracci, Jiaxin Cao, Yulong Cao, Mark Carlson, Carlos Casanova, Ting-Yun Chang, Yan Chang, Yu-Wei Chao, Prithvijit Chattopadhyay, Roshan Chaudhari, Chieh-Yun Chen, Junyu Chen, Ke Chen, Qizhi Chen, Wenkai Chen, Xiaotong Chen, Yu Chen, An-Chieh Cheng, Click Cheng, Xiu Chia, Jeana Choi, Chaeyeon Chung, Wenyan Cong, Yin Cui, Magdalena Dadela, Nalin Dadhich, Wenliang Dai, Joyjit Daw, Alperen Degirmenci, Rodrigo Vieira Del Monte, Robert Denomme, Sameer Dharur, Marco Di Lucca, Ke Ding, Wenhao Ding, Yifan Ding, Yuzhu Dong, Nicole Drumheller, Yilun Du, Aigul Dzhumamuratova, Aleksandr Efitorov, Hamid Eghbalzadeh, Naomi Eigbe, Imad El Hanafi, Hassan Eslami, Benedikt Falk, Jiaojiao Fan, Jim Fan, Amol Fasale, Sergiy Fefilatyev, Liang Feng, Francesco Ferroni, Sanja Fidler, Xiao Fu, Vikram Fugro, Prashant Gaikwad, TJ Galda, Katelyn Gao, Yihuai Gao, Wenhang Ge, Sreyan Ghosh, Arushi Goel, Vivek Goel, Akash Gokul, Rama Govindaraju, Jinwei Gu, Miguel Guerrero, Elfie Guo, Aryaman Gupta, Siddharth Gururani, Hugo Hadfield, Song Han, Ankur Handa, Zekun Hao, Mohammad Harrim, Ali Hassani, Nathan Hayes-Roth, Yufan He, Chris Helvig, Cyrus Hogg et al. (195 additional authors not shown)}, year{2026}, eprint{2606.02800}, archivePrefix{arXiv}, primaryClass{cs.CV}, url{https://arxiv.org/abs/2606.02800}, }【免费下载链接】cann-recipes-embodied-ai本项目针对具身智能业务中的典型模型、加速算法提供基于CANN平台的优化样例项目地址: https://gitcode.com/cann/cann-recipes-embodied-ai创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考