New Documentation available!
Harini Sridhar
New Documentation Available: XLA Flags Guidance and SparseCore
We are pleased to announce the release of documentation for
XLA flags - a resource many of you have frequently asked for. This guide offers a curated selection of key XLA flags designed to help you effectively navigate and utilize XLA's capabilities, with a particular focus on flags that impact runtime performance and HBM utilization.
The documentation covers:
Performance Flags: Learn about flags related to Pipelining and Asynchronous operations, and how to use latency-bound flags for all-gather, all-reduce, collective-permute, and all-to-all operations to optimize small-sized operations, especially in inference workloads.
High Bandwidth Memory Flags: Understand how to adjust HBM flags when encountering HBM "out of memory" errors during model compilation, including those for the Scheduler and Fusion.
We recommend reviewing this guide to enhance your understanding and usage of XLA. The documentation can be found at https://openxla.org/xla/flags_guidance.
SparseCore - this valuable new resource is designed to provide an in-depth look into SparseCore, helping you to use its features effectively. Whether you are looking to optimize performance for sparse workloads or better understand advanced capabilities, this guide is an essential resource.
We encourage you to explore this new guide now available at: https://openxla.org/xla/sparsecore.
- XLA team
FAQs:
Is this list of XLA flags exhaustive?
No, this documentation provides a curated selection of common and impactful XLA flags, particularly those affecting performance and HBM. The full set of XLA flags is extensive and subject to change.
Can enabling these flags negatively impact performance or stability?
While XLA flags are designed to offer control and optimization, they should be adjusted with caution. If problems arise after enabling a flag, we recommend reverting to default settings and reporting the issue.
How do I report an issue or suggest an improvement for this documentation?
Your feedback is valuable! If you run into issues, please file a bug or reach out to me.