Overview
CANN Recipes 是围绕 CANN 平台的实践集合，覆盖训练、推理、具身智能、空间智能与 HarmonyOS 端侧部署。它由 5 个仓库组成，每个仓库对应一个领域，并提供可复现的样例、性能优化说明与工程化脚本。
What You Get
| Card |
Description |
| Runnable Examples |
可直接运行的样例工程与脚本 |
| Performance Playbook |
针对 Ascend 平台的性能优化策略 |
| Deployment Patterns |
从单卡到多机的部署范式与配置参考 |
| Device-Cloud Practices |
面向真实业务的端侧推理实践 |
Repository Scope
| Area |
Focus |
| Training |
RL / SFT / multimodal training optimization |
| Inference |
LLM & multimodal inference deployment |
| Embodied Intelligence |
manipulation & policy models |
| Spatial Intelligence |
3D / spatial vision models |
| Harmony Inference |
HarmonyOS device-side ops and demos |
Documentation Goals
| Goal |
Why It Matters |
| Faster Entry |
快速找到适合你场景的样例 |
| Clear Path |
明确环境、依赖与运行路径 |
| Consistent Patterns |
形成跨仓库的统一理解与操作习惯 |
Navigation Dimensions
| Dimension |
Examples |
| User Level |
Beginner / Intermediate / Advanced |
| Scenario |
Training / Inference / Embodied / Spatial / Device-Cloud |
| Capability |
Operator / Graph Engine / Multi-Stream / Compute-Communication Fusion / Scheduling / Parallelism |
| Platform & Goal |
A2 / A3 / 310P / HarmonyOS, Low Latency / High Throughput / Long Context |