By CANN Capability

按 CANN 的能力入口查找对应实践。

Operator Fusion

Card	Description	Link
GPT-OSS Fusion	MoE/Attention 融合算子	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/gpt-oss/gpt_oss_optimization.md
HunyuanImage-3.0 Fusion	MoE 与并行优化	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/hunyuan-image-3.0/hunyuan_image_3_optimization.md
Pi0 Fusion	具身模型融合算子适配	https://gitcode.com/cann/cann-recipes-embodied-intelligence/-/blob/master/manipulation/pi0/infer_with_torch/README.md
VGGT Fusion	NPU 融合算子替换	https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/models/vggt/vggt_optimization.md

Graph Engine / Graph Mode

Card	Description	Link
TorchAir GE (RL)	训练场景图模式优化	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/llm_rl/deepseek_rl_train_optimization.md
Hunyuan3D Graph Mode	端到端图模式推理	https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/models/Hunyuan3D/Hunyuan3D_optimization.md
Pi0 Graph Mode	具身推理图模式	https://gitcode.com/cann/cann-recipes-embodied-intelligence/-/blob/master/manipulation/pi0/infer_with_torch/README.md

Multi-Stream / Compute-Communication Overlap

Card	Description	Link
DeepSeek RL Multi-Stream	MoE/MLA 多流并行	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/llm_rl/deepseek_rl_train_optimization.md
LongCat-Flash Multi-Stream	低时延推理优化	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/longcat-flash/README.md
HunyuanImage-3.0 Multi-Stream	MoE 多流并行	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/hunyuan-image-3.0/hunyuan_image_3_optimization.md

Parallelism (TP/EP/CP/DP)

Card	Description	Link
Kimi-K2-Thinking	并行策略细节	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/kimi-k2-thinking/kimi_k2_thinking_inference_guide.md
DeepSeek-V3.2-Exp	CP/EP 并行策略	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/models/deepseek-v3.2-exp/README.md
Qwen3 Long Seq RL	训练并行与调度	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/llm_rl/qwen3_235B_32k_longseq_rl_train_optimization.md

Scheduling / Load Balancing

Card	Description	Link
RL Rollout Rebalance	推理阶段负载均衡	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/features/rollout_rebalance.md
3DGS Load Balance	AscendC 负载均衡	https://gitcode.com/cann/cann-recipes-spatial-intelligence/-/blob/master/docs/algorithms/gaussian_splatting/gaussian_splatting_load_balance_optimization.md

Memory Layout / KV Cache

Card	Description	Link
KV Cache NZ	推理 KV Cache 布局优化	https://gitcode.com/cann/cann-recipes-train/-/blob/master/docs/llm_rl/deepseek_rl_train_optimization.md

Custom Operator Development

Card	Description	Link
AscendC Dev Guide	自定义算子开发与入图	https://gitcode.com/cann/cann-recipes-harmony-infer/-/blob/master/docs/ascendc_develop_guide.md
TileLang Operator	DSL 算子开发实践	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/docs/models/deepseek-v3.2-exp/deepseek_v3.2_exp_tilelang_operator_guide.md
AscendC Operator	融合算子优化实践	https://gitcode.com/cann/cann-recipes-infer/-/blob/master/docs/models/deepseek-v3.2-exp/deepseek_v3.2_exp_ascendc_operator_guide.md