Raise err if minimum Accelerate version isn't available (#22841)

* Add warning about accelerate

* Version block Accelerate

* Include parse

* Apply suggestions from code review

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

* Check partial state

* Update param

---------

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

src/transformers/training_args.py

@@ -1531,6 +1531,10 @@ class TrainingArguments:
     def _setup_devices(self) -> "torch.device":
         requires_backends(self, ["torch"])
         logger.info("PyTorch: setting up devices")
+        if not is_sagemaker_mp_enabled() and not is_accelerate_available(check_partial_state=True):
+            raise ImportError(
+                "Using the `Trainer` with `PyTorch` requires `accelerate`: Run `pip install --upgrade accelerate`"
+            )
         if self.no_cuda:
             self.distributed_state = PartialState(cpu=True)
             device = self.distributed_state.device

src/transformers/utils/import_utils.py

@@ -575,8 +575,15 @@ def is_protobuf_available():
     return importlib.util.find_spec("google.protobuf") is not None
 
 
-def is_accelerate_available():
-    return importlib.util.find_spec("accelerate") is not None
+def is_accelerate_available(check_partial_state=False):
+    accelerate_available = importlib.util.find_spec("accelerate") is not None
+    if accelerate_available:
+        if check_partial_state:
+            return version.parse(importlib_metadata.version("accelerate")) >= version.parse("0.17.0")
+        else:
+            return True
+    else:
+        return False
 
 
 def is_optimum_available():