A Chinese-language field manual for the full AI infrastructure stack

What it does

This repository collects detailed Chinese-language guides across the entire AI infrastructure chain: GPU/TPU architecture, PCIe and NVLink interconnects, CUDA and TileLang programming, Kubernetes-based GPU virtualization with HAMi and Kueue, InfiniBand/NCCL cluster operations, and even Huawei Ascend NPU porting with CANN. It reads like a syllabus someone actually finished instead of just bookmarked.

The interesting bit

Most “AI fundamentals” content stops at model architecture. This one keeps going downward—through container runtime GPU mechanics, Device Plugin source analysis, and DPU programming with NVIDIA’s DOCA framework—on the theory that the bottleneck is usually under the floorboards.

Key highlights

• Hardware coverage spans NVIDIA GB300 NVL72 rack-scale systems down to PCIe P2PDMA and GPUDirect RDMA mechanics
• GPU programming section links 200+ optimized Tensor Core/CUDA Core kernel examples and contrasts SIMT with Tile-Based (TileLang) models
• Kubernetes AI infrastructure includes practical HAMi resource isolation configs, LWS distributed inference scheduling, and JuiceFS/DeepSeek 3FS storage architecture
• Huawei NPU track covers end-to-end migration: PyTorch adapter setup, custom operators, FlashAttention porting, Qwen2.5 inference, and LoRA fine-tuning
• Cluster ops section treats nvidia-smi utilization myths and IB health checks as first-class topics, not afterthoughts

Caveats

• Entirely Chinese-language; no English translations visible in the source tree
• Some sections (DPU programming, Flex AI) appear to have shallow coverage relative to the CUDA and Kubernetes tracks
• Repository is HTML-markdown documentation, not runnable code or frameworks

Verdict

Worth a read if you’re building or operating Chinese-language AI infrastructure and need a single map of the territory. Skip it if you want copy-paste deployment templates or English prose.