Support HF LLaMA ckpt conversion

[lsy323]

Added --from_hf option in convert_checkpoint.py for HF checkpoint. Only LLaMA is supported now. Quantization conversion is not supported with HF checkpoint.

Enable converting HF llama checkpoint by

python -m convert_checkpoints --model_name=llama-2 \
    --input_checkpoint_dir=$input_ckpt_dir \
    --output_checkpoint_dir=$output_ckpt_dir \
    --from_hf=True

The guide to add support for HF checkpoint will be done in a following PR.

Only tested with HF 7B model, 70B not tested yet

[FanhaiLu1]

Thanks for support HF llama chpt conversion! Can you save hf weight names as a file in the repo?

[FanhaiLu1]

I feel [gate|down|up]_proj are more read friendly than w1, w2 and w3. @qihqi Shall we consider rename it to proj related name in default checkpoint convert?

[lsy323]

Yeah, this makes sense. Also want to note that original llama weight is using w1/2/3 https://github.com/meta-llama/llama3/blob/main/llama/model.py#L219. If we change it we need to do the name mapping for the original llama weight.

[FanhaiLu1]

Did you test the llama2-70B model?

[lsy323]

No I didn't, since I haven't set up multi host yet. But I will do that later