Update trainer.mdx class_weights example (huggingface#23787)

class_weights tensor should follow model's device

Author: amitportnoy

docs/source/en/main_classes/trainer.mdx

@@ -61,7 +61,7 @@ class CustomTrainer(Trainer):
         outputs = model(**inputs)
         logits = outputs.get("logits")
         # compute custom loss (suppose one has 3 labels with different weights)
-        loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0]))
+        loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0], device=model.device))
         loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
         return (loss, outputs) if return_outputs else loss
 ```