milti-gpu: fix inputs_embeds + position_embeds

Fixing the following errors in few models:
```
>       hidden_states = inputs_embeds + pos_embeds
E       RuntimeError: Expected all tensors to be on the same device, but found at least two devices, xpu:2 and xpu:3!
```

Fixes: huggingface#35762
Signed-off-by: Dmitry Rogozhkin <[email protected]>

src/transformers/models/gpt2/modeling_gpt2.py

@@ -818,7 +818,7 @@ def forward(
         if inputs_embeds is None:
             inputs_embeds = self.wte(input_ids)
         position_embeds = self.wpe(position_ids)
-        hidden_states = inputs_embeds + position_embeds
+        hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
 
         # Attention mask.
         _use_sdpa = self._attn_implementation == "sdpa" and output_attentions is False and head_mask is None

src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

@@ -959,7 +959,7 @@ def forward(
         if inputs_embeds is None:
             inputs_embeds = self.wte(input_ids)
         position_embeds = self.wpe(position_ids)
-        hidden_states = inputs_embeds + position_embeds
+        hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
 
         if token_type_ids is not None:
             token_type_embeds = self.wte(token_type_ids)

src/transformers/models/imagegpt/modeling_imagegpt.py

@@ -775,7 +775,7 @@ def forward(
         if inputs_embeds is None:
             inputs_embeds = self.wte(input_ids)
         position_embeds = self.wpe(position_ids)
-        hidden_states = inputs_embeds + position_embeds
+        hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
 
         if token_type_ids is not None:
             token_type_embeds = self.wte(token_type_ids)

src/transformers/models/opt/modeling_opt.py

@@ -882,7 +882,7 @@ def forward(
         if self.project_in is not None:
             inputs_embeds = self.project_in(inputs_embeds)
 
-        hidden_states = inputs_embeds + pos_embeds
+        hidden_states = inputs_embeds + pos_embeds.to(inputs_embeds.device)
 
         if self.gradient_checkpointing and self.training:
             if use_cache:

src/transformers/models/xglm/modeling_xglm.py

@@ -594,7 +594,9 @@ def forward(
                 encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
             )
 
-        hidden_states = inputs_embeds + self.embed_positions(position_ids, past_key_values_length)
+        hidden_states = inputs_embeds + self.embed_positions(position_ids, past_key_values_length).to(
+            inputs_embeds.device
+        )
         hidden_states = nn.functional.dropout(hidden_states, p=float(self.dropout), training=self.training)
 
         if self.gradient_checkpointing and self.training: