[usability] Accelerate Support

[wheresmyhair]

Description

1. LMFlow now defaults to using Accelerate (i.e, run scripts using accelerate launch ... finetune.py ...). If you prefer to use deepspeed (deepspeed ... finetune.py ...) or accelerate + deepspeed backend, please install using pip install -e '.[deepspeed]'
2. Removed/archived some less frequently used docs/scripts/modules.
   • docker
   • scripts/data_preprocess, scripts/speculative_decoding, scripts/tools, scripts/vocab_extension
   • service
   • utils
3. LISA callback and create_customized_optimizer are separated from finetuner.py
4. Add unittest for finetuner.py
5. Add calculation for # tokens and samples in the training dataset, coulde enable using --calculate_dataset_stats true
6. Fix resume from lora checkpoint. --resume_from_checkpoint path/to/your/lora/checkpoint-50

Known Issues

1. hf_encoder_decoder_model is disabled temporarily. We will have a comprehensive update for encoder-decoder model later.

Tests

scripts	Single GPU	Multiple GPUs
run_finetune.sh	✅	✅
run_finetune_with_lora.sh	✅	✅
run_finetune_with_qlora.sh	✅	✅
run_finetune_with_lisa.sh	✅	⌛
run_finetune_with_custom_optim.sh	✅	✅

[research4pan]

Need beta best before merging into main branch, since this will be a major overhaul and may have some compatibility risks for some users.

I would suggest highlighting and redirecting users to this branch in README first. Also, we can tag the current main branch first.