[trainer] `--load_best_model_at_end` silently turns of `--save_steps` settings

Splitting off from https://github.com/huggingface/transformers/pull/12477#discussion_r668326212

Currently --load_best_model_at_end silently turns off --save_steps settings when --do_eval is off (or --evaluation_strategy is set to other than "no", which otherwise automatically turns on --do_eval)

The proposal is to assert if:

--load_best_model_at_end is set and --evaluation_strategy is "no"

Reproducible test:

export BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --output_dir output_dir --adam_eps 1e-06  --do_train --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 500 --max_source_length 128 --max_target_length 128 --val_max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --predict_with_generate --sortish_sampler --source_lang en --target_lang ro --dataset_name wmt16 --dataset_config "ro-en" --source_prefix "translate English to Romanian: " --warmup_steps 50 --max_train_samples 50 --save_steps 1 

which saves checkpoints.

then adding --load_best_model_at_end stops saving those.

@sgugger.

[sgugger]

Yes, as said in that comment, I think it’s reasonable if we raise an error if --load_best_model_at_end is set and --evaluation_strategy is “no” since there is no “best model” to pick from in that case. I can do it later today if you want.

[sgugger]

Yes, this was fixed by #12786 in the end.