[Badcase]: 使用Reasoning数据微调2.0成功，但是2.5失败

[UESTCthb]

Model Series

Qwen2.5

What are the models used?

Qwen2.5-72B-Instruct

What is the scenario where the problem happened?

Qwen2.5-72B-Instruct用reasoning数据微调，模型无法收敛

Is this badcase known and can it be solved using avaiable techniques?

• I have followed the GitHub README.
• I have checked the Qwen documentation and cannot find a solution there.
• I have checked the documentation of the related framework and cannot find useful information.
• I have searched the issues and there is not a similar one.

Information about environment

GPU： H100
DATASET： KingNish/reasoning-base-20k
batch size： 1
max_length:int = 3000
batch_size:int = 1
gradient_accumulation_steps:int = 16
log_iter:int = 10
max_lr:float = 2e-3
min_lr:float = 2e-4
warmup_steps:int = 1000

Description

Steps to reproduce

This happens to Qwen2.5-72B-Instruct.
The badcase can be reproduced with the following steps:
max_length:int = 3000
batch_size:int = 1
gradient_accumulation_steps:int = 16
log_iter:int = 10
max_lr:float = 2e-3
min_lr:float = 2e-4
warmup_steps:int = 1000
使用如上参数微调reasoning 推理模型。数据集：KingNish/reasoning-base-20k。相同的代码和参数在Qwen2.0-72B-Instruct上成功，但是Qwen2.5-72B-Instruct无法收敛。backward报错。
并且对于Qwen2.0-72B-Instruct，如果学习率降低一个量级也会出现无法收敛的情况

[jklj077]

what's the error in the backward pass?