fix: fix gradient accumulate step for learning rate (huggingface#27667)

examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py

@@ -640,7 +640,7 @@ def compute_metrics(preds, labels):
 
     # Create learning rate schedule
     linear_decay_lr_schedule_fn = create_learning_rate_fn(
-        len(vectorized_datasets["train"]),
+        total_train_steps,
         training_args.warmup_steps,
         training_args.learning_rate,
     )