https://stonesjtu.github.io/ Kaiyu Shi AI infrastructure and compiler systems engineer building open, practical AI systems so everyone can enjoy AI. 2026-09-14T15:29:13+08:00 Kaiyu Shi skyisno.1@gmail.com https://stonesjtu.github.io/2026/09/09/apple-gpu-evolution/ Apple GPU Evolution 2026-09-10T17:14:00+08:00 2026-09-09T00:00:00+08:00 Apple GPU evolution from Dynamic Caching to Neural Accelerators, with an MLX Metal GEMM source walkthrough. https://stonesjtu.github.io/2026/09/04/rubin-lut-decompression/ Rubin 3-Bit LUT Decompression 2026-09-04T00:00:00+08:00 2026-09-04T00:00:00+08:00 3-bit weight storage with MXFP8 compute at 3.375 bits per weight. https://stonesjtu.github.io/2026/09/02/apple-silicon-ai-compute-gpu-vs-ane/ Apple Silicon AI compute: ANE and GPU 2026-09-10T16:04:00+08:00 2026-09-02T00:00:00+08:00 ANE and GPU hardware, software stacks, and runnable Core ML, PyTorch MPS, MLX, and Metal examples. https://stonesjtu.github.io/2026/07/03/ai-infra-edge-intelligence/ How intelligent can the edge become? - AI-Infra Overview PART-4 2026-08-09T23:40:14+08:00 2026-07-03T00:00:00+08:00 A hardware, economics, and model-quality view of how much intelligence can move into phones, AI PCs, and edge AI boxes. https://stonesjtu.github.io/2026/07/02/ai-infra-future/ Predicting the future AI infra stack - AI-Infra Overview PART-3 2026-08-10T22:25:18+08:00 2026-07-02T00:00:00+08:00 A supply-chain view of future AI infrastructure: heterogeneous computing, dedicated LLM hardware, memory locality, and interconnect. https://stonesjtu.github.io/2026/07/01/ai-infra-scaling-problem/ Why AI infra is fast-moving and challenging - AI-Infra Overview PART-2 2026-08-09T23:40:14+08:00 2026-07-01T00:00:00+08:00 Why rapidly evolving models, request volume, context, output, agents, and hardware keep moving the AI infrastructure frontier. https://stonesjtu.github.io/2026/06/30/ai-infra-and-tokenomics/ From conventional infra to agentic infra - AI-Infra Overview PART-1 2026-08-09T23:40:14+08:00 2026-06-30T00:00:00+08:00 A history of AI infrastructure through the co-evolution of model execution patterns and CPU/GPU/TPU/NPU hardware. https://stonesjtu.github.io/2025/08/14/hopper-kernel-for-mxfp4/ Hopper kernel for MXFP4 2026-09-01T15:38:52+08:00 2025-08-14T00:00:00+08:00 Notes on Triton's Hopper MXFP4 matmul path, from packed FP4 unpacking into BF16 to scale conversion with inline PTX. https://stonesjtu.github.io/2018/01/06/build-pytorch-with-nccl2/ Build PyTorch with NCCL2 2026-06-20T13:59:19+08:00 2018-01-06T00:00:00+08:00 Covers NCCL2 installation, clean PyTorch source builds, multi-GPU all-reduce tests, and network interface troubleshooting. https://stonesjtu.github.io/2018/01/05/build-pytorch-with-cuda-aware-mpi/ Build PyTorch with CUDA-aware MPI support 2026-06-20T13:59:19+08:00 2018-01-05T00:00:00+08:00 Notes on building OpenMPI with CUDA support, compiling PyTorch against MPI compilers, and checking runtime distributed support. https://stonesjtu.github.io/2018/01/04/pytorch-cuda-performance-profiling/ Trace PyTorch CUDA performance bottlenecks 2026-06-20T13:59:19+08:00 2018-01-04T00:00:00+08:00 Shows a profiling workflow from synchronized Python-level cProfile and line-profiler runs down to nvprof and kernel-level investigation. https://stonesjtu.github.io/2018/01/03/pytorch-tensor-add-explained/ Figure out what b.add_(b) has done 2026-06-20T13:59:19+08:00 2018-01-03T00:00:00+08:00 Traces PyTorch tensor addition from Python bindings through ATen, TH tensor macros, BLAS paths, OpenMP splitting, and AVX2 vector kernels. https://stonesjtu.github.io/2018/01/02/pytorch-dispatch-variable-index-select/ How PyTorch dispatches Variable.index_select 2026-06-20T13:59:19+08:00 2018-01-02T00:00:00+08:00 Explains how Tensor and Variable calls move through generated bindings, C/C++ backends, CUDA kernels, and derivative definitions. https://stonesjtu.github.io/2018/01/01/pytorch-python-to-cpp-walkthrough/ From Python to C/C++: a PyTorch walk-through 2026-06-20T13:59:19+08:00 2018-01-01T00:00:00+08:00 Walks through a compact autograd example to show how Variables, Functions, the backward graph, and the imperative engine fit together.