Skip to content

11. 延伸阅读 ​

论文 ​

奠基

  • Triton: an Intermediate Language and Compiler for Tiled Neural Network Computations(MAPL 2019,arXiv:2002.06810)——OpenAI Triton 的原始论文,tile 级抽象的起点
  • TVM: An Automated End-to-End Optimization Stack for Deep Learning(OSDI 2018,arXiv:1802.04799)——图级编译 + auto-tuning 的奠基工作
  • Ansor: Generating High-Performance Tensor Programs for Deep Learning(OSDI 2020,arXiv:2006.06762)——学习型 cost model + 层级搜索空间
  • Roofline: an insightful visual performance model(Williams et al., CACM 2009)——算术强度与"算力受限 vs 带宽受限"的理论基础
  • MLIR: Scaling Compiler Infrastructure for Domain Specific Computation(CGO 2021,arXiv:2002.11054)——dialect 体系的官方阐述

Attention kernel(编译器优化的试金石)

  • FlashAttention(arXiv:2205.14135)/ FlashAttention-2(arXiv:2307.08691)/ FlashAttention-3(arXiv:2407.08608)——手写 CUDA 路线的三次迭代
  • FlashAttention 4(2025)——用 Triton Gluon 重写并超过手写 CUDA,"DSL 反超 CUDA"的标志性案例

前沿(2025-2026)

  • Mirage: A Multi-Level Superoptimizer for Tensor Programs(OSDI 2025)——跨层级搜索的张量程序超优化器
  • MPK: A Compiler and Runtime for Mega-Kernelizing Tensor Programs(OSDI 2026)——persistent megakernel 编译 + GPU 内调度运行时
  • TileLang: A Composable Tiled Programming Model for AI Systems(arXiv:2502.02613)——基于 TVM 的 tile DSL 与其多后端设计

官方文档与工程资源 ​

  • Triton:github.com/triton-lang/triton(release notes 是跟踪 3.x 新特性的最好来源);Triton 官方 tutorials(matmul / fused-softmax / persistent-kernel)
  • Gluon:Triton 仓库内的 Gluon 示例(特别是 attention 相关)——TTGIR 级显式编程的教材
  • PyTorch 编译栈:pytorch.org/docs/stable/torch.compiler(torch.compile 权威参考);PyTorch 2.x release notes(Inductor/CuTeDSL/FlexAttention 的演进);pytorch.org/blog(FlexAttention、Helion on TPU 等深度博文)
  • cuTile:docs.nvidia.com/cuda/cutile-python(官方文档);NVIDIA Developer Blog "Simplify GPU Programming with NVIDIA CUDA Tile in Python"——60 行 Python 写 attention 的入门
  • CUTLASS:github.com/NVIDIA/cutlass(4.x release notes:Primitives API、Task Scheduling、cute_ext 管线)
  • TileLang:github.com/tile-ai/tilelang(含 LSP、pass visualizer、benchmark 仓库链接)
  • Helion:PyTorch Conference 2025 talk "Helion: A High-level DSL for Kernel Authoring"(Jason Ansel, Meta)
  • XLA / StableHLO:github.com/openxla/xla、github.com/openxla/stablehlo
  • Apache TVM:github.com/apache/tvm(Relax/TensorIR 方向)

本手册相关主题 ​

  • Triton Inference Server:撞名提醒——推理服务框架,与 OpenAI Triton 无关
  • GPU 与 CUDA:kernel 的硬件执行模型、CUDA graphs 原理
  • vLLM:torch.compile + CUDA graphs 的深度用户
  • SGLang:RadixAttention 与编译器栈的结合
  • TensorRT-LLM:"全编译"路线的代表(对比 torch.compile 的运行时编译)
  • 分布式系统:编译器与通信算子(all-reduce 重叠)的交汇

Released under CC-BY-SA-4.0 License.