Fix trainer seq2seq qa.py evaluate log and ft script (#19208)

* fix args option

* fix trainer eval log

* fix out of memory qa script

* do isort, black, flake

* fix tokenize target

* take it back.

* fix: comment

examples/pytorch/question-answering/trainer_seq2seq_qa.py

@@ -15,12 +15,14 @@
 """
 A subclass of `Trainer` specific to Question-Answering tasks
 """
+import math
+import time
 from typing import Dict, List, Optional
 
 from torch.utils.data import Dataset
 
 from transformers import Seq2SeqTrainer, is_torch_tpu_available
-from transformers.trainer_utils import PredictionOutput
+from transformers.trainer_utils import PredictionOutput, speed_metrics
 
 
 if is_torch_tpu_available(check_device=False):
@@ -59,6 +61,7 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
         # Temporarily disable metric computation, we will do it in the loop here.
         compute_metrics = self.compute_metrics
         self.compute_metrics = None
+        start_time = time.time()
         eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
         try:
             output = eval_loop(
@@ -71,6 +74,15 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
             )
         finally:
             self.compute_metrics = compute_metrics
+        total_batch_size = self.args.eval_batch_size * self.args.world_size
+        output.metrics.update(
+            speed_metrics(
+                metric_key_prefix,
+                start_time,
+                num_samples=output.num_samples,
+                num_steps=math.ceil(output.num_samples / total_batch_size),
+            )
+        )
 
         if self.post_process_function is not None and self.compute_metrics is not None:
             eval_preds = self.post_process_function(eval_examples, eval_dataset, output)
@@ -81,15 +93,15 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
                 if not key.startswith(f"{metric_key_prefix}_"):
                     metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
 
-            self.log(metrics)
-        else:
-            metrics = {}
+            output.metrics.update(metrics)
+
+        self.log(metrics)
 
         if self.args.tpu_metrics_debug or self.args.debug:
             # tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
             xm.master_print(met.metrics_report())
 
-        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, metrics)
+        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, output.metrics)
         return metrics
 
     def predict(