[deepspeed] offload + non-cpuadam optimizer exception doc (#22044)

* [deepspeed] offload + non-cpuadam optimizer exception doc

* deps

docs/source/en/main_classes/deepspeed.mdx

@@ -1293,8 +1293,17 @@ If you want to use another optimizer which is not listed above, you will have to
 }
 ```
 
-Similarly to `AdamW`, you can configure other officially supported optimizers. Just remember that may have different
-config values. e.g. for Adam you will want `weight_decay` around `0.01`.
+Similarly to `AdamW`, you can configure other officially supported optimizers. Just remember that those may have different config values. e.g. for Adam you will want `weight_decay` around `0.01`.
+
+Additionally, offload works the best when it's used with Deepspeed's CPU Adam optimizer. If you want to use a different optimizer with offload, since `deepspeed==0.8.3` you need to also add:
+
+
+```json
+{
+   "zero_force_ds_cpu_optimizer": false
+}
+```
+to the top level configuration.
 
 
 

setup.py

@@ -106,7 +106,7 @@ _deps = [
     "dataclasses",
     "datasets!=2.5.0",
     "decord==0.6.0",
-    "deepspeed>=0.6.5",
+    "deepspeed>=0.8.3",
     "dill<0.3.5",
     "evaluate>=0.2.0",
     "fairscale>0.3",

src/transformers/dependency_versions_table.py

@@ -12,7 +12,7 @@ deps = {
     "dataclasses": "dataclasses",
     "datasets": "datasets!=2.5.0",
     "decord": "decord==0.6.0",
-    "deepspeed": "deepspeed>=0.6.5",
+    "deepspeed": "deepspeed>=0.8.3",
     "dill": "dill<0.3.5",
     "evaluate": "evaluate>=0.2.0",
     "fairscale": "fairscale>0.3",