fix type annotations for arguments in training_args (#24550)

* testing

* example script

* fix typehinting

* some tests

* make test

* optional update

* Union of arguments

* does this fix the issue

* remove reports

* set default to False

* documentation change

* None support

* does not need None

* Fix typing annotations for FSDP and DeepSpeed in TrainingArguments (#24549)

* Fix typing annotations for FSDP and DeepSpeed in TrainingArguments

* Change dict to Dict

* Revert "Fix typing annotations for FSDP and DeepSpeed in TrainingArguments" (#24574)

Revert "Fix typing annotations for FSDP and DeepSpeed in TrainingArguments (#24549)"

This reverts commit c5e29d4381.

* Fix typing annotations for FSDP and DeepSpeed in TrainingArguments (#24549)

* Fix typing annotations for FSDP and DeepSpeed in TrainingArguments

* Change dict to Dict

* merge

* hacky fix

* fixup

---------

Co-authored-by: Max Ryabinin <mryabinin0@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/training_args.py

@@ -406,7 +406,7 @@ class TrainingArguments:
             When resuming training, whether or not to skip the epochs and batches to get the data loading at the same
             stage as in the previous training. If set to `True`, the training will begin faster (as that skipping step
             can take a long time) but will not yield the same results as the interrupted training would have.
-        sharded_ddp (`bool`, `str` or list of [`~trainer_utils.ShardedDDPOption`], *optional*, defaults to `False`):
+        sharded_ddp (`bool`, `str` or list of [`~trainer_utils.ShardedDDPOption`], *optional*, defaults to `''`):
             Use Sharded DDP training from [FairScale](https://github.com/facebookresearch/fairscale) (in distributed
             training only). This is an experimental feature.
 
@@ -421,7 +421,7 @@ class TrainingArguments:
 
             If a string is passed, it will be split on space. If a bool is passed, it will be converted to an empty
             list for `False` and `["simple"]` for `True`.
-        fsdp (`bool`, `str` or list of [`~trainer_utils.FSDPOption`], *optional*, defaults to `False`):
+        fsdp (`bool`, `str` or list of [`~trainer_utils.FSDPOption`], *optional*, defaults to `''`):
             Use PyTorch Distributed Parallel Training (in distributed training only).
 
             A list of options along the following:
@@ -969,7 +969,7 @@ class TrainingArguments:
             )
         },
     )
-    sharded_ddp: str = field(
+    sharded_ddp: Optional[Union[List[ShardedDDPOption], str]] = field(
         default="",
         metadata={
             "help": (
@@ -980,7 +980,7 @@ class TrainingArguments:
             ),
         },
     )
-    fsdp: str = field(
+    fsdp: Optional[Union[List[FSDPOption], str]] = field(
         default="",
         metadata={
             "help": (
@@ -1019,11 +1019,11 @@ class TrainingArguments:
             )
         },
     )
-    deepspeed: Optional[str] = field(
+    deepspeed: Optional[Union[str, Dict]] = field(
         default=None,
         metadata={
             "help": (
-                "Enable deepspeed and pass the path to deepspeed json config file (e.g. ds_config.json) or an already"
+                "Enable deepspeed and pass the path to deepspeed json config file (e.g. `ds_config.json`) or an already"
                 " loaded json file as a dict"
             )
         },