hiyouga / LLaMA-Factory

Efficiently Fine-Tune 100+ LLMs in WebUI (ACL 2024)
https://arxiv.org/abs/2403.13372
Apache License 2.0
31.01k stars 3.82k forks source link

qwen2 72b 910b lora后merge生成的权重 推理失败 #4659

Open wphtrying opened 2 months ago

wphtrying commented 2 months ago

Reminder

System Info

image

Reproduction

 ### model

model_name_or_path: ${data_dir}/qwen2_sft/merge_Qwen2-72B-lora_gaokao904 cache_dir: /cache logging_dir: ${output_dir}/log

method

stage: sft do_predict: true finetuning_type: full

dataset

dataset_dir: ${data_dir}/gaokao/ dataset: gaokao_904_test template: qwen cutoff_len: 1024

max_samples: 100

overwrite_cache: true preprocessing_num_workers: 2

output

output_dir: ${data_dir}/qwen2_infer/lora_infer overwrite_output_dir: false

eval

do_sample: false temperature: 1.0 top_p: 1.0 top_k: 50 bf16: true per_device_eval_batch_size: 1 predict_with_generate: true

ddp

ddp_timeout: 180000000 deepspeed: examples/deepspeed/ds_z3_offload_config.json

Expected behavior

No response

Others

No response

panjiacheng commented 1 month ago

求问你这个问题解决了吗?我也遇到类似的