swap batch size for gradient accumulation steps to decouple from num gpu

Update wandb_log_model on llama_65B_alpaca.yml

fix sharegpt handling from hf, don't worry about loading llama if using earlier transformers release

fix lora target module, require explicit flash attention, fix min logging steps, don't use adam8bit for int4, hash prepared datasets, support hf hub datasets

4bit quantized support (wip)

deepspeed doesn't work with flash-attn, and the gpu savings w flash attn are better than the deepspeed headaches

add llama 7b config and fiz lora_fan_in_fan_out for llama (copy pasta bug)

more logging, wandb fixes

improve prepared dataset loading, fix inference

helpful info output

various bugfixes

more fixes and prep for llama training