| set -eux | |
| LLM_RECIPES_DIR=/code/llm-recipes | |
| source $LLM_RECIPES_DIR/scripts/wmt2024/tokens.sh | |
| python /code/llm-recipes/tools/hf_inference_distrubuted.py \ | |
| --model /work/models/translation_finetuned_hf/llama2-ja-zh-continuous-pretrained-v0-dev-finetune-chunked-docs-cleaned-all-averaged-246-250 \ | |
| -i /work/wmt2024_test/LLM/split/ja-zh/wmttest2024.src.sentence_splited.with_template.ja-zh.ja.jsonl.05 \ | |
| -o /work/translation/wmt2024_test/ja-zh/llama2-top-p-0.95/split_05 \ | |
| -g 0 1 2 3 4 5 6 7 \ | |
| -b 503 \ | |
| --attn_implementation sdpa \ | |
| --dynamic_max_new_token_ratio 2.0 \ | |
| --num_return_sequences 20 \ | |
| --do_sample \ | |
| --top_p 0.95 \ | |
| --max_input_tokens 503 | |