Training Recipes

面向 LLM / 多模态训练优化的实践合集。

Repo: https://gitcode.com/cann/cann-recipes-train

Featured Recipes

Card	Level	Description	Link
Qwen2.5 RL (Starter)	Beginner	单卡 Atlas A2 入门样例，使用 verl 在数学推理数据集上训练	https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen2_5/verl_npu_demo/README.md
Qwen3 Tool Agent RL	Intermediate	端到端 agent RL 训练，启用 asyncLLM 与 agent_loop	https://gitcode.com/cann/cann-recipes-train/-/blob/master/agent_rl/qwen3_tool_agent/README.md
DeepSeek-R1 RL	Advanced	veRL + MindSpeed + vLLM-Ascend，在 Atlas A3 上实现 GRPO 高吞吐训练	https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/deepseek/README.md
Qwen3-235B-A22B RL	Advanced	2k+32k 长序列 GRPO/DAPO 训练优化	https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen3/README.md
Qwen3-32B SAM	Advanced	RL 训练场景启用 SAM 投机推理，性能提升约 10%	https://gitcode.com/cann/cann-recipes-train/-/blob/master/llm_rl/qwen3/README.md

Related Features

Card	Description	Link
SAM Speculative Decoding	SAM 无损投机推理	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/features/sam_speculative_decoding.md
Rollout Rebalance	RL 推理阶段调度与均衡	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/features/rollout_rebalance.md