This chapter is divided into eight parts; they are: • Metrics for LLM Inference • Measuring a Single Request • Warmup and Synchronization • Measuring GPU Work with CUDA Events • Measuring Memory Usage • Measuring Concurrent Requests • Multiple GPUs and...

This chapter is divided into nine parts; they are: • Reading Logits from a Model • Greedy Decoding • Temperature Sampling • Top-$k$ Sampling • Nucleus Sampling • Repetition Penalties • Beam Search • Stop Conditions • Structured Output Constraints The m...

This chapter is divided into four parts; they are: • Autoregressive Generation • Prefill and Decode • A Simple KV Cache • Memory Usage of the KV Cache A decoder-only transformer model predicts the next token from the tokens that came before it.

This article is divided into five parts; they are: • An Example of Tensor Parallelism • Setting Up Tensor Parallelism • Preparing Model for Tensor Parallelism • Train a Model with Tensor Parallelism • Combining Tensor Parallelism with FSDP Tensor paral...

This article is divided into five parts; they are: • Introduction to Fully Sharded Data Parallel • Preparing Model for FSDP Training • Training Loop with FSDP • Fine-Tuning FSDP Behavior • Checkpointing FSDP Models Sharding is a term originally used in...

This article is divided into six parts; they are: • Pipeline Parallelism Overview • Model Preparation for Pipeline Parallelism • Stage and Pipeline Schedule • Training Loop • Distributed Checkpointing • Limitations of Pipeline Parallelism Pipeline para...

This article is divided into two parts; they are: • Data Parallelism • Distributed Data Parallelism If you have multiple GPUs, you can combine them to operate as a single GPU with greater memory capacity.

This article is divided into two parts; they are: • Using `torch.

This article is divided into three parts; they are: • Floating-point Numbers • Automatic Mixed Precision Training • Gradient Checkpointing Let's get started! The default data type in PyTorch is the IEEE 754 32-bit floating-point format, also known as s...

This article is divided into two parts; they are: • What Is Perplexity and How to Compute It • Evaluate the Perplexity of a Language Model with HellaSwag Dataset Perplexity is a measure of how well a language model predicts a sample of text.

This article is divided into three parts; they are: • Training a Tokenizer with Special Tokens • Preparing the Training Data • Running the Pretraining The model architecture you will use is the same as the one created in the

This article is divided into two parts; they are: • Simple RoPE • RoPE for Long Context Length Compared to the sinusoidal position embeddings in the original Transformer paper, RoPE mutates the input tensor using a rotation matrix: $$ \begin{aligned} X...

This post is divided into five parts; they are: • Preparing the Dataset for Training • Implementing the Seq2Seq Model with LSTM • Training the Seq2Seq Model • Using the Seq2Seq Model • Improving the Seq2Seq Model In

This post is divided into three parts; they are: • Why Skip Connections are Needed in Transformers • Implementation of Skip Connections in Transformer Models • Pre-norm vs Post-norm Transformer Architectures Transformer models, like other deep learning...

This post covers three main areas: • Why Mixture of Experts is Needed in Transformers • How Mixture of Experts Works • Implementation of MoE in Transformer Models The Mixture of Experts (MoE) concept was first introduced in 1991 by

This post is divided into three parts; they are: • Why Linear Layers and Activations are Needed in Transformers • Typical Design of the Feed-Forward Network • Variations of the Activation Functions The attention layer is the core function of a transfor...

This post is divided into five parts; they are: • Why Normalization is Needed in Transformers • LayerNorm and Its Implementation • Adaptive LayerNorm • RMS Norm and Its Implementation • Using PyTorch's Built-in Normalization Normalization layers improv...

This post is divided into four parts; they are: • Why Attention Masking is Needed • Implementation of Attention Masks • Mask Creation • Using PyTorch's Built-in Attention In the

This post is divided into three parts; they are: • Low-Rank Approximation of Matrices • Multi-head Latent Attention (MLA) • PyTorch Implementation Multi-Head Attention (MHA) and Grouped-Query Attention (GQA) are the attention mechanisms used in almost ...

This post is divided into three parts; they are: • Why Attention is Needed • The Attention Operation • Multi-Head Attention (MHA) • Grouped-Query Attention (GQA) and Multi-Query Attention (MQA) Traditional neural networks struggle with long-range depen...