Support for transformers versions < 4.46.1 will soon be discontinued due to issues with incorrect gradient accumulation.

[chenzihan1]

Reminder

• I have read the README and searched the existing issues.

System Info

该LLama-Factory版本只支持transformers versions < 4.46.1，但是这个版本存在梯度计算错误的问题。是否考虑更新版本来解决该问题。

Reproduction

grpo_02: Traceback (most recent call last):
grpo_02: File "/mnt/tenant-home_speed/czh/360-LLaMA-Factory-sp/src/train.py", line 15, in
grpo_02: from llamafactory.train.tuner import run_exp
grpo_02: File "/mnt/tenant-home_speed/czh/360-LLaMA-Factory-sp/src/llamafactory/train/tuner.py", line 25, in
grpo_02: from ..hparams import get_infer_args, get_train_args
grpo_02: File "/mnt/tenant-home_speed/czh/360-LLaMA-Factory-sp/src/llamafactory/hparams/init.py", line 20, in
grpo_02: from .parser import get_eval_args, get_infer_args, get_train_args
grpo_02: File "/mnt/tenant-home_speed/czh/360-LLaMA-Factory-sp/src/llamafactory/hparams/parser.py", line 44, in
grpo_02: check_dependencies()
grpo_02: File "/mnt/tenant-home_speed/czh/360-LLaMA-Factory-sp/src/llamafactory/extras/misc.py", line 83, in check_dependencies
grpo_02: require_version("transformers>=4.41.2,<=4.46.1", "To fix: pip install transformers>=4.41.2,<=4.46.1")
grpo_02: File "/root/miniconda3/envs/lf_sp/lib/python3.10/site-packages/transformers/utils/versions.py", line 111, in require_version
grpo_02: _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
grpo_02: File "/root/miniconda3/envs/lf_sp/lib/python3.10/site-packages/transformers/utils/versions.py", line 44, in _compare_versions
grpo_02: raise ImportError(
grpo_02: ImportError: transformers>=4.41.2,<=4.46.1 is required for a normal functioning of this module, but found transformers==4.48.3.
grpo_02: To fix: pip install transformers>=4.41.2,<=4.46.1

Expected behavior

No response

Others

No response

[HaoshengZou]

可以先注释掉
require_version("transformers>=4.41.2,<=4.46.1", "To fix: pip install transformers>=4.41.2,<=4.46.1")

然后用你想用的transformer版本。

建议使用LLaMA-Factory主仓最新划定的transformer版本
https://github.com/hiyouga/LLaMA-Factory/blob/main/src/llamafactory/extras/misc.py#L91
https://github.com/hiyouga/LLaMA-Factory/blob/main/src/llamafactory/extras/misc.py#L97

我们之前也发现4.46梯度累积有bug #1 ，< 4.46的反而不算bug