qwen2vl全量微调后进行评估报错

[gsfsdv]

Description

作者您好，我全量微调的qwen2vl后，使用llamafactory-cli train examples/extras/nlg_eval/qwen.yaml进行评估

model_name_or_path: /mnt/dolphinfs/checkpoint-846

method

do_predict: true
stage: sft
finetuning_type: full

dataset

eval_dataset: text2textsftEval
dataset_dir: /mnt/dolphinfs/results
template: qwen2_vl
cutoff_len: 8192
max_samples: 2
max_new_tokens: 8192
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/qwen/
overwrite_output_dir: true

eval

per_device_eval_batch_size: 1
predict_with_generate: true
ddp_timeout: 180000000

但是有错：
File "/home/hadoop-aipnlp/.local/lib/python3.9/site-packages/transformers/models/qwen2_vl/modeling_qwen2_vl.py", line 582, in forward
attn_weights = attn_weights + causal_mask
RuntimeError: The size of tensor a (792) must match the size of tensor b (397) at non-singleton dimension 3

Pull Request

No response