Quickstart

用 5 分钟确定你的路线，并进入可运行的样例。

Step 1: Choose Your Track

Track	Description	Entry
Training	需要在 Ascend 集群上进行 RL / SFT 训练优化	recipes/train.md
Inference	需要大模型高吞吐或低时延部署	recipes/infer.md
Embodied Intelligence	需要机器人操作类模型推理或训练	recipes/embodied.md
Spatial Intelligence	需要 3D / 空间视觉模型推理或训练	recipes/spatial.md
HarmonyOS Inference	需要端侧推理和自定义算子实践	recipes/harmony.md

Step 2: Pick a Quickstart

Area	Entry	Why
LLM Inference	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/gpt-oss/README.md	部署路径清晰，适合入门
LLM Training	https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen2_5/verl_npu_demo/README.md	单卡 Atlas A2 上手
Embodied	https://gitcode.com/cann/cann-recipes-embodied-intelligence/-/blob/master/manipulation/pi0/infer_with_torch/README.md	Graph Mode + Operator Fusion
Spatial	https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/models/vggt/README.md	推理与评测脚本完整
HarmonyOS	https://gitcode.com/cann/cann-recipes-harmony-infer/-/blob/master/ops/ascendc/docs/custom-npu_sobel.md	端侧算子路径最短

Step 3: Run the Example

每个样例都包含以下最小信息：

Item	Description
Environment	环境与依赖
Model & Data	模型与数据准备
Launch	启动脚本
Results	性能与精度说明

如果你需要统一的环境清单，参考 getting-started/environment.md。