@@ -23,6 +23,7 @@ FairScale was designed with the following values in mind:
...
@@ -23,6 +23,7 @@ FairScale was designed with the following values in mind:
## What's New:
## What's New:
* November 2021 [fairscale 0.4.3 was released](https://github.com/facebookresearch/fairscale/releases/tag/v0.4.3).
* November 2021 [fairscale 0.4.3 was released](https://github.com/facebookresearch/fairscale/releases/tag/v0.4.3).
* We have experimental support for offloading params to disk when using the FSDP API for evaluation workloads.
* We have an experimental layer that fuses multiple layers together to support large vocab size trainings.
* We have an experimental layer that fuses multiple layers together to support large vocab size trainings.
* November 2021 [fairscale 0.4.2 was released](https://github.com/facebookresearch/fairscale/releases/tag/v0.4.2).
* November 2021 [fairscale 0.4.2 was released](https://github.com/facebookresearch/fairscale/releases/tag/v0.4.2).
* We have a new experimental API called the LayerwiseMemoryTracker to help track, visualize and suggest fixes for memory issues occurring during the forward/backward pass of your models.
* We have a new experimental API called the LayerwiseMemoryTracker to help track, visualize and suggest fixes for memory issues occurring during the forward/backward pass of your models.
