python -m torch.distributed.launch \
    --nproc_per_node=8 \
    run_xtreme_s.py \
    --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
    --task="fleurs-lang_id" \
    --language="all" \
    --output_dir="xtreme_s_xlsr_300m_fleurs_langid" \
    --num_train_epochs=5 \
    --per_device_train_batch_size=8 \
    --per_device_eval_batch_size=1 \
    --gradient_accumulation_steps=1 \
    --learning_rate="3e-4" \
    --warmup_steps=2000 \
    --evaluation_strategy="steps" \
    --max_duration_in_seconds=20 \
    --save_steps=1000 \
    --eval_steps=1000 \
    --logging_steps=1 \
    --layerdrop=0.0 \
    --freeze_feature_encoder \
    --gradient_checkpointing \
    --fp16 \
    --fp16_full_eval \
    --do_train \
    --do_eval \
    --do_predict \
    --metric_for_best_model="accuracy" \
    --greater_is_better=True \
    --load_best_model_at_end \
    --push_to_hub