Remove JSON config mangling for Gemma ckpt

[lsy323]

The fix to the invalid JSON config file in HF Gemma pytorch ckpt is merged. We don't need to fix the invalid json in convert script.

https://huggingface.co/google/gemma-7b-it-pytorch/discussions/2#6667fe530ec7ed4422eb070c
https://huggingface.co/google/gemma-7b-pytorch/discussions/2#6667fdbf647001c39240a47e
https://huggingface.co/google/gemma-2b-pytorch/discussions/2#6667fdab51545a8b46c3a121

Other gemma pytorch ckpts are handled by HF staff as well.

Test

export input_ckpt_dir=/mnt/disks/lsiyuan/gemma_weight/gemma-7b-pytorch-it
export output_ckpt_dir=/mnt/disks/lsiyuan/gemma_weight/gemma-7b-pytorch-it-bf16
export model_name="gemma"
export quantize_weights=False
export quantize_type="int8_per_channel"
export from_hf=False
python -m convert_checkpoints --model_name=$model_name \
    --input_checkpoint_dir=$input_ckpt_dir \
    --output_checkpoint_dir=$output_ckpt_dir \
    --quantize_weights=$quantize_weights \
    --quantize_type=$quantize_type \
    --from_hf=$from_hf

export tokenizer_path=/mnt/disks/lsiyuan/gemma_weight/gemma-7b-pytorch-it/tokenizer.model
export size="7b"
export quantize_weights=False
export quantize_activation=False
export quantize_kv_cache=False

python run_interactive.py --model_name=$model_name --size=$size --batch_size=2 --max_cache_length=2048 \
    --checkpoint_path=$output_ckpt_dir \
    --tokenizer_path=$tokenizer_path \
    --quantize_kv_cache=$quantize_kv_cache \
    --quantize_weights=$quantize_weights \
    --quantize_type=$quantize_type \

[wang2yn84]

Thank you very much for raising the issue and get it solved!