Docs
CANN Recipes
面向 CANN 平台的训练、推理与多领域实践文档集合，覆盖 Operator Fusion、Graph Engine、Multi-Stream、Parallelism、Scheduling 等能力。这里以“路径+Recipe+能力索引”的方式组织内容，方便不同用户快速找到可运行的样例与最佳实践。
文档正文
面向 CANN 平台的训练、推理与多领域实践文档集合，覆盖 Operator Fusion、Graph Engine、Multi-Stream、Parallelism、Scheduling 等能力。这里以“路径+Recipe+能力索引”的方式组织内容，方便不同用户快速找到可运行的样例与最佳实践。
快速开始
| 入口 | 说明 | Link |
|---|---|---|
| 入门路线 | 了解整体结构与推荐路径 | getting-started/overview.md |
| 快速上手 | 选择场景并进入第一个 Recipe | getting-started/quickstart.md |
| 环境准备 | 通用依赖与平台说明 | getting-started/environment.md |
Quickstarts
| Area | Recipe | Why | Link |
|---|---|---|---|
| LLM Inference | GPT-OSS | 体量较小、部署路径清晰 | https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/gpt-oss/README.md |
| LLM Training | Qwen2.5 RL | 单卡 Atlas A2 入门样例 | https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen2_5/verl_npu_demo/README.md |
| Embodied | Pi0 (Torch) | Graph Mode + Operator Fusion | https://gitcode.com/cann/cann-recipes-embodied-intelligence/-/blob/master/manipulation/pi0/infer_with_torch/README.md |
| Spatial | VGGT | 评测与推理脚本完整 | https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/models/vggt/README.md |
| HarmonyOS | SobelCustom | 端侧算子路径最短 | https://gitcode.com/cann/cann-recipes-harmony-infer/-/blob/master/ops/ascendc/docs/custom-npu_sobel.md |
按场景进入
| 场景 | 说明 | 入口 |
|---|---|---|
| LLM / Multimodal Training | 训练优化与 RL/SFT 实践 | recipes/train.md |
| LLM / Multimodal Inference | 低时延与高吞吐部署 | recipes/infer.md |
| Embodied Intelligence | 具身智能推理与训练 | recipes/embodied.md |
| Spatial Intelligence | 3D 与空间视觉 | recipes/spatial.md |
| HarmonyOS / Device-Cloud | 端云协同与 AscendC | recipes/harmony.md |
入口索引
| 维度 | 入口 |
|---|---|
| 用户层级 | paths/by-level.md |
| 场景 | paths/by-scenario.md |
| 能力 | paths/by-capability.md |
| 平台 | paths/by-platform.md |
| 目标 | paths/by-goal.md |
CANN 能力地图
| Capability | 说明 | 入口 |
|---|---|---|
| Operator Fusion | 融合算子实践与替换策略 | paths/by-capability.md |
| Graph Engine / Graph Mode | 图模式与整图优化 | paths/by-capability.md |
| Multi-Stream | 通信与计算重叠 | paths/by-capability.md |
| Parallelism | TP/EP/CP/DP 并行策略 | paths/by-capability.md |
| Scheduling | 负载均衡与调度策略 | paths/by-capability.md |
| Custom Operator | AscendC / TileLang 实践 | paths/by-capability.md |
精选 Recipe
| Level | Recipe | 说明 | Link |
|---|---|---|---|
| Beginner | GPT-OSS | 推理入门 | https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/gpt-oss/README.md |
| Beginner | Qwen2.5 RL | 训练入门 | https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen2_5/verl_npu_demo/README.md |
| Intermediate | LongCat-Flash | 低时延推理 | https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/longcat-flash/README.md |
| Intermediate | Hunyuan3D | Graph Mode + Operator Fusion | https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/models/Hunyuan3D/README.md |
| Advanced | DeepSeek-V3.2-Exp | 高吞吐推理 | https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/deepseek-v3.2-exp/README.md |
| Advanced | Qwen3-235B RL | 长序列训练 | https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen3/README.md |
仓库地图
| Repo | Focus | Primary Audience |
|---|---|---|
| cann-recipes-train | LLM / multimodal training optimization | Training engineers, algorithm devs |
| cann-recipes-infer | LLM / multimodal inference & deployment | Inference engineers, infra teams |
| cann-recipes-embodied-intelligence | Embodied manipulation models | Robotics / embodied AI developers |
| cann-recipes-spatial-intelligence | 3D / spatial intelligence models | 3D CV / spatial AI devs |
| cann-recipes-harmony-infer | HarmonyOS device-side inference | HarmonyOS app & NPU devs |
参考与贡献
| 入口 | Link |
|---|---|
| FAQ | reference/faq.md |
| 贡献指南 | contributing/index.md |