doc(cpu-offload): modify README to cover XPU

Signed-off-by: dbyoung18 <yang5.yang@intel.com>

Author: dbyoung18

torchao/prototype/low_bit_optim/README.md

@@ -80,7 +80,7 @@ All of our low-bit optimizers mentioned above also support `bf16_stochastic_roun
 
 ## Optimizer CPU offload
 
-This folder also implements optimizer CPU offload (i.e. ZeRO-Offload) for single GPU training. Only CUDA is supported. For multi-GPU training, you can use FSDP's built-in CPU offload.
+This folder also implements optimizer CPU offload (i.e. ZeRO-Offload) for single GPU training. Only CUDA and XPU is supported. For multi-GPU training, you can use FSDP's built-in CPU offload.
 
 ```python
 import torch
@@ -97,7 +97,7 @@ optim = CPUOffloadOptimizer(model.parameters(), torch.optim.AdamW, offload_gradi
 
 This will reduce GPU memory usage by optimizer state size, and additionally gradient size if `offload_gradients=True`. `CPUOffloadOptimizer` can wrap any base optimizer.
 
-For saving and loading `CPUOffloadOptimizer`, it is important that you load model's weights BEFORE creating the optimizer, since we create a CPU copy of the parameters inside `CPUOffloadOptimizer.__init__()`. (TODO: we might want to have a method to synchronize CUDA and CPU params in either direction CPU->CUDA and CUDA->CPU, in case they are out of sync.)
+For saving and loading `CPUOffloadOptimizer`, it is important that you load model's weights BEFORE creating the optimizer, since we create a CPU copy of the parameters inside `CPUOffloadOptimizer.__init__()`. (TODO: we might want to have a method to synchronize GPU and CPU params in either direction CPU->GPU and GPU->CPU, in case they are out of sync.)
 
 ```python
 ckpt = torch.load("checkpoint.pth")