scripts/decode/en-ja/llama2/top_p_inference_2.sh

set -eux
LLM_RECIPES_DIR=/code/llm-recipes
source $LLM_RECIPES_DIR/scripts/wmt2024/tokens.sh

for i in `seq 7 9`; do
    GPU_ID=$((i-5))
    python /code/llm-recipes/tools/hf_inference.py \
	   --model /work/models/additiona_trained_hf/llama2-en-ja-continuous-pretrained-v0-dev-finetune-chunked-docs-all-averaged-841-845 \
	   -i /work/wmt2024_test/LLM/split/en-ja/wmttest2024.src.sentence_splited.with_template.en-ja.en.jsonl.0${i} \
	   -o /work/translation/wmt24_test/en-ja/llama2-top-p-0.95/split_0${i} \
	   -g ${GPU_ID} \
	   -b 158 \
	   --attn_implementation sdpa \
	   --dynamic_max_new_token_ratio 3.0 \
	   --num_return_sequences 50 \
	   --do_sample \
	   --top_p 0.95 \
	   --max_input_tokens 158 &
done
wait