PegasusX add _no_split_modules (huggingface#25933)

* no_split_modules

* no_split_modules

* inputs_embeds+pos same device

* update _no_split_modules

* update _no_split_modules

src/transformers/models/pegasus_x/modeling_pegasus_x.py

@@ -769,6 +769,7 @@ class PegasusXPreTrainedModel(PreTrainedModel):
     config_class = PegasusXConfig
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
+    _no_split_modules = [r"PegasusXEncoderLayer", r"PegasusXDecoderLayer"]
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -1299,6 +1300,8 @@ def forward(
         # embed positions
         positions = self.embed_positions(inputs_embeds, past_key_values_length)
 
+        positions = positions.to(inputs_embeds.device)
+
         hidden_states = inputs_embeds + positions
 
         hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)