remove redunant call to eval and train (#2372)

trl/trainer/callbacks.py

@@ -53,7 +53,7 @@ def _generate_completions(
     batch_size: int = 1,
 ) -> List[str]:
     """
-    Generates completions for a list of pre-formatted prompts.
+    Generates completions for a list of pre-formatted prompts from the given model.
 
     Args:
         prompts (List[str]): A list of input prompts for which completions are to be generated.
@@ -68,7 +68,6 @@ def _generate_completions(
     """
     completions = []
     with unwrap_model_for_generation(model, accelerator) as unwrapped_model:
-        unwrapped_model.eval()
         for idx in range(0, len(prompts), batch_size):
             batch = prompts[idx : idx + batch_size]
             tokenized_batch = tokenizer(batch, return_tensors="pt", padding=True, truncation=True).to(model.device)
@@ -81,7 +80,6 @@ def _generate_completions(
                 generation = generation[len(prompt) :]
                 completion = tokenizer.decode(generation, skip_special_tokens=True)
                 completions.append(completion)
-        unwrapped_model.train()
     return completions