LLM Systems 17: System for MoE Models 0. Recap Pipeline Parallelism split by layers (horizonal split) eliminate the bubbles (idle) interleaving forward/backward Tensor Parallelism split the matrix computation over multiple G 2026-10-31 ML System #CMU 11868 #LLM System #Distributed Training #MoE
LLM Systems 16: Model Parallel Training 0.Recap Data Parallel via All Reduce Distributed Data Parallel Training gradient bucketing overlay backward and AllReduce communication 1. Model Parallel Training Motivation: 模型规模呈指数增长,单 2026-10-31 ML System #CMU 11868 #LLM System #Distributed Training
LLM Systems Assignment 4: Transformer CUDA Acceleration 本次 homework 主要内容是:实现 softmax 与 LayerNorm kernel fusion 形式的 fwd 与 bwd,完成 CUDA kernel 编写;接入 Python 层;接入 Transformer 模型并验证训练加速。 下面对本次 homework 实现时一些注意点进行介绍。 1. Softmax 1.1 forward 课程已经提供了 block_dim 2026-10-17 ML System #CMU 11868 #LLM System #kernel
DDP: PyTorch Distributed-Experiences on Accelerating Data Parallel Training PyTorch Distributed: Experiences on Accelerating Data Parallel Training 代码:PyTorch DDP arXiv 提交日期:28 Jun 2020 (发表于 VLDB 2020) 作者: 摘要: 本文介绍了 PyTorch 分布式数据并行(Distributed Data Parallel, D 2026-10-17 ML System #CMU 11868 #LLM System #DDP
LightSeq2: Accelerated Training for Transformer-based Models on GPUs LightSeq2: Accelerated Training for Transformer-based Models on GPUs arXiv 提交日期:12 Oct 2021 GitHub:https://github.com/bytedance/lightseq 作者: 摘要 基于 Transformer 的神经网络模型已被广泛应用于许多人工智能领域。训练这 2026-10-17 ML System #CMU 11868 #LLM System #kernel
LightSeq: A High Performance Inference Library for Transformers LightSeq: A High Performance Inference Library for Transformers arXiv 提交日期:23 Oct 2020 GitHub:https://github.com/bytedance/lightseq 作者: 摘要 Transformer、BERT 及其变体在自然语言处理领域取得了巨大的成功。由于 Tran 2026-10-17 ML System #CMU 11868 #LLM System #kernel
LLM Systems 15: Distributed Data Parallel Training 0. Recap Overall idea: partition the data, distribute the forward/backward Parameter Server server to update and distribute parameters, worker to get local grad NCCL Multi-GPU communicatio 2026-10-17 ML System #CMU 11868 #LLM System #Distributed Training
LLM Systems 14: Distributed Training 0. Recap 之前课程主线: LLM 简介 CUDA 编程基础 深度学习框架(自动求导) Transformer/Pre-trained LLM LLM 前后两个辅助组件:Tokenizer 与 Decoding Transformer 加速:Kernel Fusion & Mix-precision & Memory Management Optimiza 2026-10-17 ML System #CMU 11868 #LLM System #Distributed Training
LLM Systems 10 & 11: Accelerating Transformer Training and Inference 0. Recap Transformer Architecture: 1. Accelerated GPU Computation for Transformer Training Overview: 对应下面将要介绍的 4 个优化手段: Kernel Fusion:计算图级别优化,通过算子融合减少计算图节点。 Reduce synchronization:算子内优化 2026-10-17 ML System #CMU 11868 #LLM System #kernel
LLM Systems 09: Decoding: Sampling, Beam Search and Speculative Decoding 0. Recap Subword tokenization: Byte-Pair-Encoding iteratively merging most frequent pairs of tokens Information-theoretic vocabulary (VOLT) solving entropy constrained optimal transport pr 2026-10-17 ML System #CMU 11868 #LLM System #Decoding