Disable CPU Offload as default for DeepSpeed (#6262)

* Change default for CPU offload to false for best throughput/memory efficiency

* Add changelog

* default

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

CHANGELOG.md

@@ -20,6 +20,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Changed the order of `backward`, `step`, `zero_grad` to `zero_grad`, `backward`, `step` ([#6147](https://github.com/PyTorchLightning/pytorch-lightning/pull/6147))
 
 
+- Changed default for DeepSpeed CPU Offload to False, due to prohibitively slow speeds at smaller scale ([#6262](https://github.com/PyTorchLightning/pytorch-lightning/pull/6262))
+
+
 - Renamed `pytorch_lightning.callbacks.swa` to `pytorch_lightning.callbacks.stochastic_weight_avg` ([#6259](https://github.com/PyTorchLightning/pytorch-lightning/pull/6259))
 
 

pytorch_lightning/plugins/training_type/deepspeed.py

@@ -66,7 +66,7 @@ def __init__(
         self,
         zero_optimization: bool = True,
         stage: int = 2,
-        cpu_offload: bool = True,
+        cpu_offload: bool = False,
         contiguous_gradients: bool = True,
         overlap_comm: bool = True,
         allgather_partitions: bool = True,
@@ -104,7 +104,7 @@ def __init__(
             stage: Different stages of the ZeRO Optimizer. 0 is disabled,
                 1 is optimizer state partitioning, 2 is optimizer+gradient state partitioning (default: 2)
 
-            cpu_offload: Enable offloading optimizer memory and computation to CPU (default: True)
+            cpu_offload: Enable offloading optimizer memory and computation to CPU
 
             contiguous_gradients: Copies gradients to a continuous buffer as they are produced.
                 Avoids memory fragmentation during backwards. Useful when training large models. (default: True)