Tune in virtually to the OpenXLA Dev Lab tomorrow!

[James Rubin]

Hello OpenXLA Community,

The OpenXLA Dev Lab kicks off tomorrow, Thursday, November 14th! Although in-person registration has been closed for awhile now, you can still catch some of the action virtually.  

We'll be offering the following sessions over Google Meet:

1/ Keynote: Get the download on 2024 Highlights and our 2025 Roadmap from Robert Hundt, Distinguished Engineer. Use this Google Meet link.

2/ Tutorials: See session descriptions and Google Meet links below.

Tutorial Name	Tutorial Description	Google Meet
Building a PJRT Plugin for a Custom Accelerator	This tutorial outlines a typical PJRT framework/plugin interaction, showcasing necessary APIs. It also examines a simple plugin using the StableHLO Reference Interpreter to evaluate programs from OpenXLA frameworks (like JAX or PyTorch/XLA), serving as a template for new plugin development.	Google Meet
StableHLO Composites: Preserving Deep Learning Ops from Frameworks and Compilers	This tutorial introduces StableHLO and HLO composite operations, demonstrating how they improve OpenXLA compiler performance without sacrificing portability. It includes a practical example of getting a composite from JAX to an OpenXLA backend.	Google Meet
Partitioning Workloads with Shardy	Shardy, a new MLIR-based partitioning system, improves upon GSPMD and PartIR. In this tutorial, we introduce a novel sharding representation, propagation algorithm, and enhanced debugging capabilities.	Google Meet
XLA GPU MLIR Codgen	New emitters in XLA:GPU landed earlier this year. They solve exponential code bloat issues, improve performance and simplify debugging. This tutorial will give an example how an HLO Fusion is optimized and compiled end to end.	Google Meet
Host Off-loading	Overview of new APIs in JAX that allow to express heterogeneous programs: buffers can live on device or on host, single StableHLO program can represent computations on device or on host. How to use them. How they are implemented in XLA. What's coming next.	Google Meet

If you encounter any technical issues or have trouble finding the right session, please email openxla-dev-...@google.com and we'll respond as soon as possible.

Best,

James

[James Rubin]

Hey Everyone,

I forgot to include the time slots for the sessions mentioned above, so here they are:

1/ Keynote: 9:00 to 9:15AM. Use this Google Meet link.

2/ Tutorials: Each tutorial is offered twice. Round 1 from 9:15 to 10:30AM and Round 2 from 10:40 to 11:55AM. The same link is used for both rounds of each tutorial. You can find all this information including links, schedule, etc. in this Agenda.

Tutorial Name	Tutorial Description	Google Meet
Building a PJRT Plugin for a Custom Accelerator	This tutorial outlines a typical PJRT framework/plugin interaction, showcasing necessary APIs. It also examines a simple plugin using the StableHLO Reference Interpreter to evaluate programs from OpenXLA frameworks (like JAX or PyTorch/XLA), serving as a template for new plugin development.	Google Meet
StableHLO Composites: Preserving Deep Learning Ops from Frameworks and Compilers	This tutorial introduces StableHLO and HLO composite operations, demonstrating how they improve OpenXLA compiler performance without sacrificing portability. It includes a practical example of getting a composite from JAX to an OpenXLA backend.	Google Meet
Partitioning Workloads with Shardy	Shardy, a new MLIR-based partitioning system, improves upon GSPMD and PartIR. In this tutorial, we introduce a novel sharding representation, propagation algorithm, and enhanced debugging capabilities.	Google Meet
XLA GPU MLIR Codgen	New emitters in XLA:GPU landed earlier this year. They solve exponential code bloat issues, improve performance and simplify debugging. This tutorial will give an example how an HLO Fusion is optimized and compiled end to end.	Google Meet
Host Off-loading	Overview of new APIs in JAX that allow to express heterogeneous programs: buffers can live on device or on host, single StableHLO program can represent computations on device or on host. How to use them. How they are implemented in XLA. What's coming next.	Google Meet

If you encounter any technical issues or have trouble finding the right session, please email openxla-dev-...@google.com and we'll respond as soon as possible.

Best,

James