LLM Systems 06: Transformer 0. Recap Learning parameters of an NN needs gradient calculation Computation Graph to perform computation: topological traversal along the DAG Auto Differentiation building backward compu 2026-10-05 ML System #CMU 11868 #LLM System #Transformer
LLM Systems 05: Deep Learning Framework and Auto Differentiation 0. Recap Tiling for efficient matrix computation Use shared memory to reduce global memory access Roofline Model & Computational Intensity Coalesced memory access Understanding the ch 2026-10-05 ML System #CMU 11868 #LLM System
LLM Systems Assignment 1: CUDA Programming 本次作业主要涉及两个文件: llmsys_hw1/minitorch/cuda_kernel_ops.py:Python 侧调用 kernel 编译成的 .so 接口 llmsys_hw1/src/combine.cu:CUDA kernel 实现 完成本次作业的主要预备知识要求是掌握基本的 CUDA 编程能力(如:按照课程 readings 要求完成 PMPP 这本书前 6 2026-09-19 ML System #CMU 11868 #LLM System #CUDA
PMPP Chapter 06: Performance considerations 本章主要内容 present the off-chip memory (DRAM) architecture and discuss related performance considerations such as memory coalescing and memory latency hiding. discuss an important type of optimizati 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
PMPP Chapter 05: Memory architecture and data locality 本章主要关注 GPU 的 on-chip memory,重点是利用 shared memory 减少 global memory 访问,提升 kernel 性能。 Importance of memory access efficiency Compute to Global Memory Access Ratio(算术强度 / 计算强度) 定义: 指程序某区域内,每 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
PMPP Chapter 04: Compute architecture and scheduling Architecture of a modern GPU GPU 计算单元由若干个 SM 组成(通常几十到一百多) 每个 SM 有若干个处理单元,称为 CUDA Cores(简称 cores) Memory(on-chip memory) Example:A100 GPU,108 SMs,每个 SM 64 个 cores global memory(off-chip 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
PMPP Chapter 03: Multidimensional grids and data 多维 grid & block gridDim gridDim.x 范围:[1, 2^{31} -1] gridDim.y & gridDim.z 范围:[1, 2^{16} -1] blockDim 最多 1024 线程(blockDim.x blockDim.y blockDim.z <= 1024) blockIdx threadId 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
PMPP Chapter 02: Heterogeneous data parallel computing 以 RGB 图片 -> 灰度图片转换为例,说明 Data Parallel 如何运行并行计算: CUDA 编程步骤: 分配 device global memory 拷贝 host memory 内容到 device memory launch kernel 在 device 进行计算 从 device 拷贝计算结果到 host 释放 device 内存 注: 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
PMPP Chapter 01: Introduction CPU vs. GPU 2003 年摩尔定律失效后,单核心 CPU 的频率/性能增长有限,新一代 CPU 的单核性能往往提升很小。面对这一问题 ,半导体产业产生了两条解决路径: 多核(multicore)CPU 多线程(multi-thread)GPU 在浮点运算性能(FLOPs)上,GPU 已经远远超过 CPU。这是因为二者设计哲学的区别: CPU 设计目的是优化 2026-09-19 ML System #CMU 11868 #LLM System #CUDA #PMPP
LLM Systems 04: GPU Acceleration 0. Recap Basic GPU CUDA operations memory allocation data movement creating threads and running on SMs specifying number of threads and number of blocks in a grid referring to data in GP 2026-09-19 ML System #CMU 11868 #LLM System