Advanced JAX: Performance, Optimization, and Scale
Chapter 1: Advanced JAX Transformations and Control Flow
Review of Core Transformations: jit, grad, vmap
Mastering lax.scan for Sequential Operations
Conditional Execution with lax.cond
Looping with lax.while_loop
Combining Control Flow and Transformations
Advanced Masking Techniques
Understanding Closures and JAX Staging
Practical: Implementing Complex Recurrent Logic
Chapter 2: Optimizing JAX Code for Performance
Profiling JAX Code on CPU, GPU, and TPU
Understanding JAX Computation Graphs (jaxpr)
The Role of XLA Compilation
Memory Layout and Its Impact on Performance
Avoiding Recompilation
Fusion and Operator Optimization
Asynchronous Dispatch
Practice: Optimizing a Numerical Computation
Chapter 3: Distributed Computing with JAX
Introduction to Parallelism Concepts
Device Management in JAX
Single-Program Multiple-Data (SPMD) with pmap
Implementing Data Parallelism using pmap
Collective Communication Primitives (psum, pmean, etc.)
Handling Axis Names in pmap
Nested pmap and Advanced Partitioning
Introduction to Multi-Host Programming
Practice: Distributed Data-Parallel Training
Chapter 4: Advanced Automatic Differentiation Techniques
Review of Forward- and Reverse-Mode Autodiff
Jacobian-Vector Products (JVPs) with jax.jvp
Vector-Jacobian Products (VJPs) with jax.vjp
Higher-Order Derivatives
Computing Full Jacobians and Hessians
Custom Differentiation Rules with jax.custom_vjp
Custom Differentiation Rules with jax.custom_jvp
Differentiation through Control Flow Primitives
Handling Non-Differentiable Functions
Practice: Implementing a Custom Gradient
Chapter 5: Interoperability and Custom Operations
Integrating JAX with NumPy
Zero-Copy Data Sharing with DLPack
Calling External CPU/GPU Code with jax.experimental.host_callback
Using jax.pure_callback for Side-Effect Free Calls
Introduction to JAX Primitives
Defining Custom Primitives
Implementing Abstract Evaluation Rules
Implementing Lowering Rules for Backends (CPU/GPU/TPU)
Defining Differentiation Rules for Custom Primitives
Practice: Integrating a C++ Function
Chapter 6: Large-Scale Model Training Techniques
Overview of Challenges in Large Model Training
Introduction to JAX Ecosystem Libraries (Flax, Haiku)
Managing Model Parameters and State
Combining pmap with Training Frameworks
Gradient Accumulation
Gradient Checkpointing (Re-materialization)
Mixed Precision Training
Model Parallelism Strategies
Optimization Algorithms for Large Scale
Practice: Implementing Gradient Checkpointing
Gradient Checkpointing (Re-materialization)
Was this section helpful?
- Training Deep Nets with Sublinear Memory Cost, Tianqi Chen, Bing Xu, Chiyuan Zhang, Carlos Guestrin, 2016 arXiv preprint arXiv:1604.06174 DOI: 10.48550/arXiv.1604.06174 - This paper introduced gradient checkpointing (also known as activation checkpointing or re-materialization) to reduce memory consumption during deep neural network training.
- jax.checkpoint, JAX developers, 2024 - Official documentation for the
jax.checkpoint(aliased asjax.remat) transformation, providing practical usage details and examples for memory optimization in JAX. - The Reversible Residual Network: Backpropagation Without Storing Activations, Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse, 2017 Advances in Neural Information Processing Systems 30 (NIPS 2017) (Curran Associates Inc.) - This paper introduces Reversible Residual Networks, an architectural design that further optimizes memory by allowing perfect reconstruction of activations during the backward pass, extending the principles of recomputation.
- ZeRO: Memory Optimizations Toward Training Trillion Parameter Models, Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He, 2020 SC20: International Conference for High Performance Computing, Networking, Storage and Analysis (IEEE) DOI: 10.1109/SC41405.2020.00024 - This paper presents ZeRO, a comprehensive set of memory optimization techniques for large-scale distributed model training, which integrates activation checkpointing within its framework.