Update flags gpu_performance_tips.md

This commit is contained in:
Abhinav Goel 2024-06-25 08:53:43 -07:00 committed by GitHub
parent a4c92a454b
commit 7b9636f63c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -66,10 +66,6 @@ training on Nvidia GPUs](https://github.com/NVIDIA/JAX-Toolbox/blob/main/rosetta
### Communication flags
* **--xla_gpu_enable_async_collectives** This flag enables the collective ops
such as `AllReduce`, `AllGather`, `ReduceScatter` and `CollectivePermute` to
be asynchronous. Asynchronous communication can overlap cross-core
communication with computation. The default value is False.
* **--xla_gpu_enable_latency_hiding_scheduler** This flag enables latency hiding
schedulers to overlap asynchronous communication with computation efficiently.
The default value is False.