Make clearer about zero_init requirements

src/transformers/modeling_utils.py

@@ -42,7 +42,11 @@
 from .configuration_utils import PretrainedConfig
 from .dynamic_module_utils import custom_object_save
 from .generation import GenerationConfig, GenerationMixin
-from .integrations import PeftAdapterMixin, deepspeed_config, is_deepspeed_zero3_enabled
+from .integrations import (
+    PeftAdapterMixin,
+    deepspeed_config,
+    is_deepspeed_zero3_enabled,
+)
 from .pytorch_utils import (  # noqa: F401
     Conv1D,
     apply_chunking_to_forward,

src/transformers/training_args.py

@@ -503,6 +503,11 @@ class TrainingArguments:
             evolve in the future. The value is either the location of DeepSpeed json config file (e.g.,
             `ds_config.json`) or an already loaded json file as a `dict`"
 
+            <Tip warning={true}>
+                If enabling any Zero-init, make sure that your model is not initialized until
+                *after* initializing the `TrainingArguments`, else it will not be applied.
+            </Tip>
+
         accelerator_config (`str`, `dict`, or `AcceleratorConfig`, *optional*):
             Config to be used with the internal `Accelerator` implementation. The value is either a location of
             accelerator json config file (e.g., `accelerator_config.json`), an already loaded json file as `dict`,