Fix the non-convergence in DSV3 post-pretrain

[zhangbo9674]

Before submitting

• Lint code. If there are lint issues, please format the code first.

# Install and register `pre-commit` in the project folder pip install pre-commit && pre-commit install # Process previous code files separately pre-commit run --file XXXX.py 

• Add test cases into tests folder. If there are codecov issues, please add tests cases first.

PR types

Bug fixes

PR changes

Models

Description

修复 DSV3 post-pretrain 热启不收敛问题，主要修改包括：

• 修复 Attention 组网代码中 rope_scaling 未正确应用，该问题会导致初始 grad norm 偏高：见 modeling.py
• 修复 FastCrossEntropyFunction 未采用 FP32 计算以及未正确开始 numeric_stable，该问题对训练后期 grad norm 的稳定性影响很高：见 modeling.py
• 修复 MoeGate 计算 seq_aux_loss 时未正确对 gates 做归一化，该问题会导致热启 loss 上扬：见 moe_gate.py
• 修复Moe专家梯度未正确裁剪问题，该问题会导致整体 grad norm 偏高：见 moe_hybrid_parallel_optimizer.py 及 trainer_callback.py
• 根据论文正确配置参数：num_nextn_predict_lambda=0.1，aux_loss_free_gamma=0.0

修复后收敛效果如下:

[paddle-bot]

Thanks for your contribution!